betaFeedback

WorksComputer Science

Agents on the Patch Path: Mining AI-Associated Security Remediation in GitHub Security Advisories

Abstract

Autonomous coding agents are increasingly involved in software development, but their participation in real-world security remediation remains poorly understood. We investigate observable evidence of coding-agent involvement in commits referenced by reviewed GitHub Security Advisories (GHSAs), characterize the associated vulnerabilities, and examine review timing in relation to advisory disclosure routes. We develop a local-Git mining pipeline that identifies agent evidence through known agent accounts, explicit commit-message attribution, repository instruction files, and agent configuration artifacts. Linking these findings to GHSA-referenced commits yields 469 valid candidate advisories across 189 repositories, including 448 with direct attribution and 21 supported only by repository-level evidence. Claude is the most frequently observed tool family, followed by Copilot and Cursor. We further compare review timing against other advisories and historical baselines using an exact route–timing decomposition. Relative to advisories published in the same months, candidate advisories exhibit a modestly longer mean review lag, primarily associated with within-route timing rather than disclosure-route composition. Compared with a historical baseline, they exhibit substantially shorter mean review lags, with both route composition and within-route timing contributing to the difference. These findings demonstrate that coding-agent involvement is already observable across hundreds of security advisories and repositories, including high- and critical-severity vulnerabilities. By distinguishing direct attribution from repository-level adoption evidence and separating route-composition effects from within-route timing differences, we provide an empirical characterization of agent-associated remediation in the GHSA review pipeline.

Introduction

AI coding agents are moving from code completion toward repository-level work: they implement features, repair defects, generate tests, respond to review comments, and create pull requests. Large-scale datasets such as AIDev [20] show that these agents already operate across tens of thousands of public repositories and leave persistent traces in software-development artifacts. Security remediation is a particularly consequential setting for such participation. A successful patch must not merely compile; it must remove the vulnerability without introducing regressions, accurately describe the affected versions, and move through disclosure channels quickly enough to support downstream users [27].

GitHub Security Advisories (GHSAs) [12] provide an unusually useful observation point. A reviewed advisory includes structured severity and weakness metadata, affected-version information, timestamps, and references that often identify remediation commits. A recent prior work [26] characterized the GHSA review pipeline and found two dominant routes: a fast path associated with GitHub Repository Advisories (GRAs) and a slower path in which information first appears in the National Vulnerability Database (NVD). That work treated patches and advisories as human-produced objects. However, the rapid adoption of coding agents creates a new question: are AI agents visible on the path from a security fix to a reviewed advisory?

Answering this question is harder than searching for tool names. A commit message may explicitly credit Claude, Copilot, Cursor, Codex, or another agent; an exact bot identity may be present in the author or committer field; or a repository may add an AGENTS.md, CLAUDE.md, or tool-specific configuration directory. However, these evidence signals do not have identical semantics. Explicit commit attribution is evidence that an agent participated in producing that commit. An instruction file is evidence of repository-level readiness or adoption, but does not prove that an agent authored the security-relevant lines. Moreover, Git metadata is self-declared and names can collide with human identities.

Therefore, we study candidate AI-associated remediation rather than claiming verified machine authorship. Our study first mines public repositories associated with reviewed advisories, using a hardened local-Git adapter and a narrow set of rules to detect coding agents. It then links findings to commit URLs in advisory references. The resulting dataset enables three complementary analyses. First, we quantify direct attribution and weaker evidence of adoption of coding agents (RQ1). Next, we characterize the security issues connected to the candidate commits (RQ2). Finally, we relate the candidates to advisory-source and review-time behavior (RQ3). These analyses correspond to the following research questions.

RQ1. To what extent do commits referenced by reviewed GHSAs carry observable evidence of AI coding agents, which evidence classes and tool families account for it, and how has it grown over time?

RQ2. What is the severity, exploitability, and weakness profile of the advisories linked to candidate AI-associated remediations, and does it concentrate on particular vulnerability classes?

RQ3. Through which disclosure routes do candidate AI-associated remediations enter the GHSA review pipeline, and does their review latency differ from that of other advisories, published in the same months or in an earlier period, once route composition is separated from within-route timing?

Our main contributions are:

  • a reproducible, local-Git method for detecting four classes of agent evidence in repositories associated with security advisories;

  • a linked dataset of 469 valid candidate advisories after removal of 29 advisories affected by human-name collisions, including a strict separation between 448 direct-attribution candidates and 21 artifact-only candidates;

  • an empirical characterization of tools, severity, CWE categories, GRA usage, and review latency in this emerging population; and

  • an exact analytical decomposition, instantiated with 12,951 same-month and 10,403 historical advisories, showing that the candidates’ review lag does not differ clearly from that of same-month advisories, and that the lower mean relative to the historical cohort is mostly route composition and calendar change, not faster review within a route.

The remainder of this paper is organized as follows. § 2 reviews the GHSA review pipeline, prior work on coding agents in real repositories, and research on security patches and disclosure latency. § 3 describes the study population, the local-Git mining procedure, the detection rules, the linkage of findings to advisories, and the validation steps. § 4 answers the three research questions. § 5 develops the route–timing decomposition that separates route composition from within-route timing. § 6 interprets the evidence and its implications for maintainers and platforms, § 7 discusses threats to validity, § 9 describes the replication package, and § 8 concludes the work.

Background and Related Work

GitHub Security Advisories

The GitHub Advisory Database aggregates vulnerability records from GitHub-sourced security advisories (GHSAs), the National Vulnerability Database (NVD), and ecosystem-specific sources. Reviewed advisories are validated and mapped to supported package ecosystems, allowing them to support services such as Dependabot alerts [15 a]. Advisories expose identifiers, publication and review timestamps, severity, CVSS information, affected packages and version ranges, CWE labels, credits, and external references. These records therefore provide a useful basis for studying both vulnerability characteristics and the processes through which vulnerabilities are disclosed and reviewed.

Prior work has examined security advisories from several complementary perspectives. Studies of vulnerability reporters and disclosure practices investigate the roles of contributors, maintainers, and reporting platforms in vulnerability management [2, 7]. Other work has examined how maintainers review and resolve vulnerability reports and how disclosure practices are reflected in advisory metadata [6, 5, 8]. Research has also investigated mechanisms for reporting security vulnerabilities through GitHub repositories (Kancharoendee et al. 2025). Together, these studies establish the importance of understanding not only vulnerabilities themselves but also the people, platforms, and workflows involved in their disclosure.

Segal et al. analyzed 288,604 advisories spanning 2019–2025, including 23,563 reviewed advisories, and characterized the GHSA review pipeline [26]. They found that security advisories (GHSAs) sourced from GRAs (GitHub Repository Advisories) reached review more quickly than those sourced from the NVD. Following the June 2022 NVD backfill1, publication-to-review delays became short for both source groups, whereas the difference in patch-to-review latency remained substantial. Their model characterized these patterns through two paths: a direct path for GRA-first advisories and an NVD-first path involving an additional stochastic delay. While this prior work also examined GHSAs, it did not mine repository histories for coding-agent evidence, link such evidence to advisory commits, characterize AI-associated remediation, or provide the route–timing decomposition introduced in our work. Thus, our study complements the pipeline-level analysis performed by Segal et al. [26] by changing the unit of interest. Rather than focusing exclusively on advisory sources and review latency, we begin with commits referenced by security advisories and investigate whether repository history contains explicit evidence of coding agent involvement. This perspective connects the disclosure process to the development activity underlying the security fixes.

AI Coding Agents and Software Security

Recent coding agents can generate and modify code, interact with development tools, and contribute to repository workflows. Their growing adoption has motivated empirical studies of both the security properties of generated code and the characteristics of agent contributions in real-world projects.

One line of research investigates the security of AI-generated code. Pearce et al. examined the security of code contributions generated by GitHub Copilot [23], while Fu et al. conducted an empirical study of security weaknesses in Copilot-generated code in GitHub projects [11]. These studies motivate continued scrutiny of AI-generated software, particularly when generated code is incorporated into production repositories. However, the presence of potential security weaknesses in generated code is distinct from evidence that a coding agent participated in a particular vulnerability fix. Establishing that connection requires examining development artifacts and their provenance.

A complementary line of research studies coding agents as participants in software development. Li et al. introduced the AIDev dataset to characterize agentic coding activity in GitHub pull requests, including contributions from tools such as Copilot, Devin, Claude Code, Cursor, and Codex [20]. More recently, Siddiq et al. investigated the security of agentic pull requests on GitHub, bringing the security implications of agent participation in repository workflows into focus [27]. This body of work moves beyond synthetic coding benchmarks to examine agent activity in real development settings.

Closer to the vulnerability-disclosure setting, Rooijendijk et al. compared how humans, bots, and autonomous agents mention vulnerability identifiers (CVE, CWE, GHSA, and others) in GitHub pull requests, using AIDev and an augmented set of pull requests from the same repositories [25]. They found that bots account for most mentions, mainly in dependency-update and audit descriptions, whereas human and agent mentions are rarer but appear in more places, such as commit messages and titles, where they support fixes and discussion. Their unit of analysis is the identifier mention in pull-request text.

Our approach is complementary to these studies in both its sampling strategy and its unit of analysis. Rather than beginning with a catalog of agent pull requests and subsequently examining their properties, we select GHSAs and the commits they reference. We then investigate whether the associated repository history contains explicit agent evidence. Depending on the available artifacts, this can include direct commit attribution, attribution retained in merged or squashed commits, Copilot Autofix markers, and repository-level instruction or configuration files.

These forms of evidence are not equivalent. Direct attribution can provide evidence of an agent’s participation in a particular change, whereas repository-level configuration artifacts establish a weaker connection to the repository’s development environment. Accordingly, identifying an agent-associated repository is not sufficient to conclude that an agent produced a particular security patch. By distinguishing evidence types, our study seeks to characterize the observable relationship between coding agents and security advisories without treating agent adoption as proof of agent authorship.

Security Patches and Vulnerability Disclosure Latency

Security patches differ from ordinary software changes in their complexity, review requirements, and coordination with releases [19]. Research on vulnerability management has examined vulnerability disclosure, patch availability, dependency exposure, and downstream adoption. For example, Arora et al. investigated software vendors’ patch-release behavior in relation to vulnerability disclosure [4]; Ponta et al. studied the detection, assessment, and mitigation of vulnerabilities in open-source dependencies [24]; and Decan et al. analyzed the impact of vulnerabilities in the npm dependency network [10]. Related work has examined the security risks associated with dependency ecosystems [29].

The timing of security patches and disclosures is crucial because publicly available patches may reveal vulnerability details before all affected users have become aware of, or adopted, a fix [28]. Research on vulnerability disclosure management has also examined the factors influencing disclosure and remediation in open-source projects [21]. These studies motivate treating the publication, patching, and review of an advisory as distinct events rather than as a single measure of vulnerability-handling speed.

Automation can influence several stages of this process. Research on automated dependency updates has investigated how tools such as Dependabot affect developer workflows and dependency-update behavior [22, 16]. More broadly, studies of GitHub Actions have examined how developers use automation to coordinate repository workflows [18]. Prior work on vulnerability databases and security-event data provides additional context for the consistency and timeliness of vulnerability information across sources [9, 3]. These studies are relevant to the present work because the path through which an advisory enters the GHSA database can affect when it becomes visible and when it is reviewed.

Coding agents introduce additional possible mechanisms. Agent participation could help developers produce patches, tests, or disclosure materials more quickly. Conversely, maintainers may need additional review or revision cycles for agent-associated security changes. Existing work on generated-code security and agentic pull requests motivates both possibilities, but does not by itself establish how agent involvement relates to the review latency of security advisories.

This study therefore distinguishes disclosure route from review timing and examines agent evidence in the commits referenced by advisories. Its focus is not to assume that agent involvement accelerates or delays security remediation, but to investigate whether explicit evidence of such involvement is associated with observable patterns in the advisory review process.

Study Methodology

Figure 1 shows an overview of our methodology, which is elaborated in details in the next subsections.

Overview of our study’s methodology

Figure 1. Overview of our study’s methodology.

Retrieving Reviewed Advisories

We collected every reviewed advisory in the GitHub Advisory Database through GitHub’s public API on September 23, 2026 obtaining a total of 35,774 advisories. We restrict the study to reviewed advisories because only these have been curated by GitHub: they are mapped to package ecosystems and affected versions, drive Dependabot alerts, and carry the review timestamp that RQ3 requires. For each advisory, we stored the complete record returned by the API. Our analyses use the source code location and references for repository linkage (§ 3.2); the severity, CVSS, EPSS, and CWE fields for the security profile (RQ2); and the publication, GitHub review, and NVD publication timestamps, together with the repository-advisory URL that identifies the disclosure route, for the timing analysis (RQ3).

Figure 2 shows the monthly publication volume of the reviewed advisories, which span October 2017 to September 2026 (the last month is partial). The isolated spike in 2022 reflects GitHub’s automated backfill of historical NVD advisories into GHSA, as explained in § 2. The volume of advisories also increases sharply in 2026: the 2026 portion of the snapshot contains 10,492 advisories, compared with 3,376–3,693 per year in 2023–2025. The growth is broad rather than driven by a single project; openclaw/openclaw, the largest contributor, accounts for 629 of these advisories. Since most candidate advisories were also published in 2026 (§ 4), calendar time is a potential confounder in any comparison with earlier advisories (§ 7).

Reviewed advisories published per month in the study population

Figure 2. Reviewed advisories published per month in the study population (n=35,774n = 35{,}774). September 2026 covers only the days before the collection date. The spike around June 2022 corresponds to the NVD backfill.

Linking Advisories to GitHub Repositories

We then identified the repositories associated with each advisory, verified their availability on GitHub, and collected their metadata. Specifically, repositories were identified primarily from the advisory’s source code location field. When this field was absent or did not resolve to a GitHub repository, we took the owner and repository from GitHub commit URLs in the advisory’s external references.

After analyzing the advisories’ metadata, we were able to locate the corresponding GitHub repository for 31,343 out of 35,774 advisories, resulting in a total of 7,739 distinct repositories; the remaining 4,431 GHSAs have neither a valid GitHub source code location nor a commit reference to a single GitHub repository. It is important to note that, at repository metadata collection time, 92 of these 7,739 repositories were no longer publicly available. Excluding the 117 advisories associated with them leaves our dataset with a total of 31,226 advisories linked to the remaining 7,647 repositories, which belong to 5,019 distinct owners (3,125 organizations and 1,894 individual users).

Figure 3 shows how advisories are distributed across the repositories identified through a valid source code location. The distribution is strongly right-skewed: most repositories have a single advisory, whereas a small tail of projects, led by openclaw/openclaw (629 advisories) and tensorflow/tensorflow (421), contributes many. Because of this concentration, our analyses deduplicate at the advisory level and use repository-cluster resampling when estimating uncertainty (§ 5.4).

Distribution of reviewed advisories per GitHub repository

Figure 3. Distribution of reviewed advisories per GitHub repository, for the repositories identified through a valid source code location. Advisories without a valid GitHub location are excluded

Mining repository histories for agent evidence

After obtaining advisory metadata and finding their corresponding GitHub repo, in this stage, we mine each of these repositories histories. Our goal in this step is to searches each repository’s version history for traces of AI coding-agent use, and § 3.5 later links the resulting findings to the commits that advisories cite. We were able to fully mine 7,643 out of the 7,647 available repositories. The remaining four repositories 2 presented problems during collection, and we could not obtain their data. Our results therefore do not cover these repositories or the 318 advisories (0.9% of the population) associated with them, 297 of which belong to liferay/liferay-portal.

For each of the 7,643 repositories associated with 30,908 advisories, we performed the following:

  1. Acquisition: We cloned each repository with its full commit history and pinned the head of its default branch. Thus, all subsequent search queries operate on the history reachable from that commit.

  1. Broad candidate search: In this step, we performed a keyword-based search over the commit’s metadata in order to identify candidate commits that might carry agent evidence. Specifically, each repository was queried with a filtered git log command over the pinned history, using fixed lists of terms and paths (Appendix § 10). The search keywords were matched as literal, case-insensitive substrings, without regular expressions or learned classifiers. The search query was deliberately designed to favor recall and covered four types of commit metadata: (a) names and e-mail addresses in the author and committer fields; (b) authorship and co-authorship references in commit messages, including Co-authored-by trailers; (c) additions or renames of agent instruction files at the repository root (e.g., AGENTS.md, CLAUDE.md); and (d) additions or modifications of agent configuration files (e.g., files under .claude, .cursor, or .codex).

  1. Deduplication: In this last step, we remove duplicated commits that were matched independently across the four searches.

It is important to note that since the keyword search was broad, and favored recall over precision, in the next step (§ 3.4) we employ deterministic rule-based detection rules to discard false positives from this list of candidate commits. This resulted in a total of 240,205 distinct commits.

Rule-based Detection of Commits that are False Positives

The candidate commits previously collected may include false positives for many reasons. For example, Git metadata is not cryptographically authenticated: a contributor can write an arbitrary co-author trailer, and a tool-like name can belong to a human. Thus, we treat every signal as a declaration and use “direct attribution” rather than “verified authorship”. Consequently, a candidate commit is only retained in our study if it is confirmed by at least one of the four deterministic, precision-oriented rules shown in Table 1. Two rules read the commit message and identity fields; the other two read the repository paths that the commit changes. The complete catalog of recognized names, handles, and paths is given in Appendix § 10.

RuleCategoryEvidence boundaryInterpretation in this study
Message signalDirect attributionExplicit attribution in the complete commit message: a well-formed co-author trailer naming a recognized agent, an explicit generation phrase, or a distinctive Copilot workflow phraseDeclared agent participation in that commit
Known bot identityDirect attributionAuthor or committer name exactly equal to an allowlisted AI-bot handle; co-author trailers excludedDeclared agent participation in that commit
Instruction fileAdoption artifactAddition or rename of a recognized, case-sensitive agent-instruction file at the repository root, such as AGENTS.md or CLAUDE.mdThe commit introduced agent-facing instructions; not commit authorship
Agent configurationAdoption artifactAddition or rename of a recognized root configuration file, or addition of a file below a recognized top-level directory such as .claude, .cursor, or .codexThe commit configured the repository for an agent tool; not tool execution or commit authorship

Table 1. Active detection rules and the claims they can support.

Commit-message attribution. The message rule accepts three kinds of signal. The first is a well-formed, line-anchored Co-authored-by trailer whose name is an allowlisted bot handle or contains a recognized tool name as a whole token, or whose address is one of the canonical Copilot, Anthropic, or OpenAI addresses. The second is an explicit generation phrase such as “generated by GitHub Copilot” or “AI-generated.” The third is a distinctive workflow form such as a standalone “Apply Copilot suggestion” line or “Copilot Autofix.”

Known bot identity. The identity rule requires the author or committer name to equal, case-insensitively, an allowlisted agent account such as copilot-swe-agent[bot] or devin-ai-integration[bot]. It does not verify the account against GitHub, and it ignores co-author trailers: a human-authored commit that lists an AI co-author is handled by the message rule rather than classified as bot-authored.

Adoption artifacts. The two path rules detect the introduction of agent-facing files: an instruction file at the repository root, such as AGENTS.md or CLAUDE.md, or project-level configuration, either a recognized root file (e.g., .cursorrules) or any file below a recognized top-level directory (e.g., .claude or .cursor). Additions and renames into these paths count, names are matched exactly and case-sensitively, and file contents are not inspected. A path finding therefore marks the commit that introduced an artifact, not the repository’s state at later commits. This distinction matters when findings are linked to advisories (§ 3.5).

Visibility of attribution. Attribution is observable only if it survives the project’s merge workflow. When a pull request is squashed or rebased, the commit that reaches the default branch carries only the metadata and message that the merge produces; depending on how that message is composed, co-author trailers from the original commits may be preserved or lost. The rules see only the commits that reach the pinned history, so the corpus reflects attribution that survives these workflows.

At the end of this step, we retained a total of 240,205 commits.

Linking evidence findings to security advisories

The previous steps produced evidence findings at the commit level, independently of any advisory. This linking step ties those findings to changes directly related to resolving a vulnerability. We therefore linked evidence and advisories through the commits that the advisories themselves cite, rather than through repository membership alone: a repository that uses an agent somewhere in its history is not, by that fact, evidence that an agent took part in a security fix.

Advisory references are the natural anchor for this association: when a fix is public, a reviewed GHSA commonly lists the URL of the remediation commit. For each advisory, we extracted the repository and commit identifier from every GitHub reference whose path contains a /commit/ segment, and we linked the advisory to every finding recorded for that repository and commit. Two constraints keep the join strict. Only full 40- or 64-character commit identifiers are used, because abbreviated identifiers can be ambiguous. The repository must match as well as the identifier, because forks share commit objects with their upstream, so an identifier-only join could attach a finding from one repository to an advisory about another. Advisories without commit references, or whose referenced commits have no findings, are outside the candidate set.

The result is a many-to-many relation: an advisory can cite several commits (for example, a fix and its backports), a commit can be cited by several advisories, and one commit can carry several findings. We therefore report counts at each level of aggregation—linked rows, advisory–commit pairs, distinct commits, and distinct advisories and repositories (Table 2)—and advisory-level summaries use one record per GHSA identifier.

MeasureCount
Advisory references to GitHub commits25,956
with parseable repository and full identifier25,601
Advisory–finding links (repository and identifier match)534
Distinct advisory–commit pairs525
Distinct linked commits440
Valid candidate advisories469
Direct attribution448
Artifact only21
Distinct repositories189

Table 2. Candidate-corpus construction

An advisory is classified as direct attribution if at least one of its linked findings comes from the message or identity rule, and as artifact only if all of its linked findings come from the path rules. The second class is narrower than its name suggests. Because a path finding marks the commit that introduced an artifact, an advisory is artifact-only when the cited commit itself added an instruction or configuration file; a fix made in a repository that already contained AGENTS.md produces no artifact link. Artifact-only advisories therefore capture agent configuration arriving in the same change as the remediation, not repository-level adoption in general. Tool-family counts are multi-label, because the findings linked to one advisory can credit more than one agent.

Validation of trace links between commits and advisories

The advisories linked to candidate commits were manually validated by two authors of this paper, who checked whether the evidence matched in each linked commit supports the signal assigned by the detection rule. After validation, 469 out of the 498 linked advisories were accepted. All 29 rejections were human-name collisions, i.e., the linked commits carried co-author trailers naming one of two human contributors whose first name is Devin, which the message rule had matched as the Devin agent. One collision affected a single advisory; the other, a contributor to a single repository, affected 32 advisories. In 28 of them the trailer was the only evidence, and we removed those advisories before classification (§ 4). In the remaining four, the same commits also carried an explicit “AI-assisted” statement; we disregarded the trailer and retained the advisories on the basis of that statement alone. The case shows that even distinctive tool names can collide with personal names, and that a single prolific contributor can multiply one collision across many advisories, so tool name matches in co-author trailers require validation rather than being accepted as attribution by construction.

Study Dataset Overview

Table 2 presents the linkage counts in our corpus used in our study. The collected advisories cite 25,956 GitHub commit URLs, of which 25,601 carry a parseable repository and full commit identifier. The strict repository-and-identifier join yields 534 advisory–finding links. A further 11 identifier matches were rejected because the cited repository differed from the mined one, mostly renamed or transferred repositories and fixes cited from personal forks. The links form 525 advisory–commit pairs, because one commit can carry more than one finding, and refer to 440 distinct commits and 498 advisories across 190 repositories. Validation removed 29 advisories whose only evidence was a co-author trailer naming a human contributor called Devin (§ 3.6), leaving 469 valid candidate advisories across 189 repositories.

Empirical Results

RQ1: Prevalence, evidence classes, tool families, and growth

We found that, of the 469 valid candidate advisories, 448 (95.5%) of them (spanning 183 repositories) contain at least one direct-attribution signal, while 21 (4.5%), from eight repositories, are supported only by an instruction or configuration artifact. At the advisory–commit level, the corresponding counts are 473 direct-attribution pairs and 21 artifact-only pairs. Direct attribution comes almost entirely from commit messages: the message rule supports 447 advisories and the bot-identity rule four. This asymmetry matters: the principal empirical result concerns direct public attribution, whereas the artifact-only subset is evidence that the repository had begun configuring an agent-oriented workflow.

Tool-family detection is multi-label and is computed over direct-attribution evidence. Claude appears in 306 of the 448 direct-attribution advisories (68.3%), Copilot in 57, Cursor in 42, Codex in 12, Gemini in four, Devin in three, CodeRabbit in two, and Aider and Sourcery in one each. A further 31 advisories carry only a generic phrase such as “AI-assisted” that names no tool. These counts should not be interpreted as market share. Tools differ in how consistently they emit trailers, and the rule catalog intentionally favors signals with distinctive names and canonical identities.

Figure 4 compares the monthly candidate counts with all advisories published in the same months. Of the 469 valid advisories, two were published in 2024, 36 in 2025, and 431 in 2026 (through September). This growth does not merely follow the volume of advisories. The monthly total also rose, from at most 422 in 2025 to between 1,366 and 1,704 from March to July 2026, but the candidates’ share of all advisories rose faster: from 0.05% in 2024 to 1.0% in 2025 and 4.1% in 2026. Monthly, the share stayed at or below 1.1% until November 2025, ranged from 1.8% to 3.7% in the first half of 2026, and reached 6.4% to 9.9% from July to September 2026. The isolated peak of December 2025 (5.1%) is mostly artifact-only evidence (11 of 19 candidates). The concentration is consistent with recent agent adoption, but it also creates a strong calendar-time confound for any comparison with historical advisories.

Monthly candidate advisories by evidence class and all reviewed advisories published in the same month

Figure 4. Monthly candidate advisories by evidence class and all reviewed advisories published in the same month, on a logarithmic scale. Direct attribution is shown with solid circles and artifact-only evidence with dashed squares; months without candidates in a class have no marker. September 2026 covers only the days before the collection date

Key Finding: Public histories provide direct agent-attribution evidence for 448 reviewed security advisories across 183 repositories. Repository artifacts identify a further 21 advisories, but those observations must not be presented as agent-authored security fixes.

Severity, exploitability, and weakness profile

Table 3 profiles the 469 valid candidates against a month-standardized reference: the 13,199 other reviewed advisories published in the 19 months that contain candidates, reweighted so that each month has the weight it has among the candidates. Because 92% of the candidates were published in 2026, this reference keeps time changes in the advisory population from being read as a property of the candidates.

MeasureCountCandidatesReference
Critical severity4910.4%10.9%
High severity20042.6%39.4%
Medium severity19341.2%41.5%
Low severity275.8%8.2%
CVSS v3 observed35074.6%72.4%
CVSS v4 only11223.9%26.5%
EPSS observed36978.7%82.1%
CWE-918 (server-side request forgery)5211.1%5.4%
CWE-22 (path traversal)5010.7%7.4%
CWE-863 (incorrect authorization)326.8%6.0%
CWE-862 (missing authorization)296.2%5.4%
CWE-94 (code injection)255.3%3.5%
CWE-306 (missing authentication)194.1%1.8%
CWE-79 (cross-site scripting)183.8%7.6%
GRA-linked38682.3%79.6%
Reviewed within 24 hours of first disclosure7716.5%20.9%

Table 3. Security profile of the valid candidate advisories (n=469n = 469) and of the month-standardized reference (n=13,199n = 13{,}199). CWE shares are multi-label. The last row is computed over the 468 candidates and the 12,951 reference advisories that are on the two routes and have a source time no later than the review (§ 4.3)

Severity closely follows the reference. High-severity advisories are the largest class (200, 42.6%), followed by medium (193, 41.2%), critical (49, 10.4%), and low (27, 5.8%); the reference shares are 39.4%, 41.5%, 10.9%, and 8.2%. Nearly all candidates carry a CVSS base score (462, 98.5%): 350 have a CVSS v3 score, with a median of 7.4 (reference 7.1) and a mean of 6.97, and a further 112 have only a CVSS v4 score. EPSS values are available for 369 candidates; their median is 0.0035 (reference 0.0034), and only two exceed 0.1, although the maximum is 0.82. The same contrast between high severity and low exploitation probability holds in the reference, so it reflects the advisory population rather than a trait of the candidates. The share of GRA-linked advisories (82.3% versus 79.6%) is also close to the reference. The share reviewed within 24 hours of first disclosure, measured as in § 4.3, is somewhat lower for the candidates (16.5% of 468 versus 20.9%); § 4.3 examines disclosure routes and review timing.

The weakness profile departs more from the reference. The candidates span 139 distinct CWE identifiers. The most frequent are CWE-918 (server-side request forgery, 52 advisories; 11.1% versus 5.4% in the reference), CWE-22 (path traversal, 50; 10.7% versus 7.4%), CWE-863 (incorrect authorization, 32; 6.8% versus 6.0%), CWE-862 (missing authorization, 29; 6.2% versus 5.4%), and CWE-94 (code injection, 25; 5.3% versus 3.5%). CWE-306 (missing authentication for a critical function) is more than twice as frequent as in the reference (4.1% versus 1.8%), whereas cross-site scripting (CWE-79), one of the most frequent weaknesses in the reference, is half as frequent (3.8% versus 7.6%). These differences are not driven by a single project: the SSRF advisories come from 28 repositories and the path-traversal advisories from 40, and the largest repository contributes 5.8% of all candidates (the five largest, 18.8%). They may nonetheless reflect the kinds of projects that adopted coding agents early rather than the agents themselves; we did not classify project domains. Given that an advisory can contain multiple CWE identifiers, these counts are multi-label.

Since the linkage follows the commits that advisories cite, this profile describes the vulnerabilities whose remediation carries public agent attribution; it does not measure the security quality of agent-written code. Advisories usually cite the fix itself, but their references can also include backports, tests, or release commits. We did not classify the role of each cited commit, so an attribution may concern a change adjacent to the fix rather than the fix itself.

Key Finding: Approximately half of the valid candidates are high or critical severity (53.1%, versus 50.3% in the month-standardized reference), and their severity, score, and exploitability profile is close to that of other advisories published in the same months. Their weakness profile differs: server-side request forgery, path traversal, code injection, and missing authentication are over-represented, and cross-site scripting is under-represented.

Disclosure route and review timing

Of the 469 valid candidates, 386 are GRA-linked, 82 are NVD-first, and one follows neither route. We set the singleton aside, leaving 468 candidates on the two substantive paths, 82.5% of them GRA-linked. We measure review timing using the review lag: the time, in days, between the publication of an advisory through the source associated with its route into the GHSA and its review by GitHub. For GRA-linked advisories, this timestamp is the publication date of the repository advisory; for NVD-first advisories, it is the NVD publication date. These timestamps identify the publication associated with each advisory’s route into GHSA and do not necessarily represent the vulnerability’s earliest public disclosure. The same definition applies to every group (§ 5.1). We compare the candidates with two reference groups drawn from the same snapshot, both excluding candidates and advisories that lack a source time or whose source time follows the review (12 contemporaneous and 72 historical advisories). The contemporaneous reference contains the 12,951 advisories published in the months that contain candidates, month-standardized as in Section 4.2. The historical reference contains the 10,403 advisories published from July 2022, after the NVD backfill, to August 2025, the period covered by our earlier study ([26]).

Route composition depends on the reference. Against the historical reference, candidates are far more concentrated in the GRA path (82.5% versus 44.3%). Against advisories published in the same months, the difference nearly vanishes (82.5% versus 80.1%): by 2026, when most candidates were published, GRA had become the dominant route for all reviewed advisories. The historical contrast therefore reflects a platform-wide shift over time rather than a property of AI-associated remediation.

Within each route, the candidates resemble the contemporaneous reference and differ sharply from the historical one (Figure 5). On the GRA-linked route the median lag is 7.4 days for the candidates, 6.0 days in the contemporaneous reference, and 0.4 days in the historical one; the shares reviewed within one day are 17.4%, 21.2%, and 69.9%. On the NVD-first route the medians are 12.0, 7.4, and 1.1 days, and the shares within one day are 12.2%, 19.8%, and 47.5%. A candidate advisory is reviewed sooner than a contemporaneous one in 45.8% of GRA-linked pairs (95% interval 40.6%–51.0%) and in 42.5% of NVD-first pairs (34.6%–51.1%), so neither difference is clear. Against the historical reference the same probabilities are 16.3% (13.1%–20.4%) and 25.5% (20.1%–33.2%). The historical reference thus describes an earlier, faster period. Its NVD-first route nevertheless has by far the heaviest extreme tail: the 95th percentile is 338.0 days, compared with 59.8 days for the candidates and 63.6 days for the contemporaneous reference, while on the GRA-linked route the historical 95th percentile (8.2 days) is far below the candidates’ (73.1 days) and the contemporaneous reference’s (69.1 days).

Empirical CDFs for GRA-linked and NVD-first routes; legend: Historical reference, Contemporaneous reference, AI-associated candidates; panels (a) GRA-linked and (b) NVD-first; x-axis: Days from source publication to review (+1 minute; logarithmic scale); y-axis: Empirical CDF

Figure 5. Empirical CDFs of the review lag from source publication to GitHub review on the GRA-linked route (a, n=386n = 386 candidates, 9,578 contemporaneous, 4,612 historical) and the NVD-first route (b, n=82n = 82 candidates, 3,373 contemporaneous, 5,791 historical). The contemporaneous reference is month-standardized. One minute is added only to display very short lags on the logarithmic axis; the reference curves are plotted at 0.5-percentage-point resolution.

The lags in Figure 5 aggregate a period in which the review queue changed. Figure 6 plots, for the advisories published in 2026, the date of GitHub’s review against the date of the source publication. Until about April, the points lie on the diagonal: advisories were reviewed almost as soon as they were disclosed. From May onward, on both routes, review dates fall progressively behind publication dates. Reviews proceed approximately in publication order (first in, first out), so the points bend away from the diagonal along a single curve, and they occur in bursts: a horizontal streak is one review date that covers advisories published over several weeks. The candidates lie on the same curve as the other advisories and inside the same bursts; neither route shows a separate fast lane or a separate slow lane for them. § 5.6 returns to this calendar trend.

Date of GitHub’s review against the date of the source publication for advisories published in 2026

Figure 6. Date of GitHub’s review against the date of the source publication for advisories published in 2026 on the GRA-linked route (a) and the NVD-first route (b). The vertical distance to the dashed diagonal is the review lag. Panel (a) shows 358 AI-associated candidates and 7,982 other advisories; panel (b) shows 73 candidates and 1,976 other advisories. Candidates are drawn over the other advisories.

Because the lag of an advisory depends mainly on when it joined this queue, and it changed quickly within a month (for non-candidate GRA-linked advisories the median rose from 0.4 days in January to 5.7 days in May and 29.3 days in July; § 5.6), a month-standardized reference leaves part of the calendar confound in place. We therefore repeat the comparison with a week-standardized reference. It is built like the contemporaneous reference, but within each route and with each source week in place of each month: the non-candidates of a route are reweighted so that every source week has the share it has among that route’s candidates. All 386 GRA-linked and all 82 NVD-first candidates have at least ten reference advisories in their week. On the GRA-linked route the candidates’ median lag is 7.4 days against 5.3 days for this reference, 17.4% of the candidates are reviewed within one day against 23.3%, and a candidate is reviewed sooner than a reference advisory in 43.5% of pairs (95% interval 38.4%–47.9%). The gap is modest, but unlike the month-standardized comparison its interval excludes one half: matched on the week, GRA-linked candidates are slower than their peers, not faster. On the NVD-first route the medians are 12.0 and 8.5 days, the shares within one day are 12.2% and 10.5%, and the probability is 50.7% (44.2%–55.4%), so the candidates are indistinguishable from their peers. These comparisons use only advisories that had been reviewed at collection.

The overall mean review lag reflects both the proportions of advisories following each route and the mean lag within each route. Advisories using coding agents average 21.95 days across the two routes, compared with 19.45 days in the contemporaneous reference and 35.23 days in the historical reference. These means are also sensitive to long lags, particularly in the historical group.

Section 5 decomposes the differences from both reference groups into route-composition and within-route components.

Key Finding: Most candidates enter through the GRA path (82.5%), but so do other advisories published in the same months (80.1%). The large gap relative to the historical cohort (44.3%) reflects the platform’s shift toward repository advisories over time. Review lags lengthened across the whole platform in 2026 as reviews fell behind publications, and the candidates sit in the same queue as other advisories. On the NVD-first route candidates are not clearly faster or slower than contemporaneous advisories; on the GRA-linked route they are not faster, and they are modestly slower than advisories published in the same week. They look slower than historical ones because review lags were far shorter in the historical period.

A Route–Timing Decomposition

Review latency is jointly determined by where an advisory enters the disclosure pipeline and how long it takes to progress within that pipeline. These mechanisms are substantively different. Agent-associated remediation may be more likely to use GitHub’s repository-advisory workflow without changing review speed within either route. Conversely, automation may change a route-specific handoff while leaving route choice unchanged.

Accounting for the two considered paths, the AI-associated candidates have a mean review lag of 21.95 days, compared with 19.45 days in the contemporaneous reference and 35.23 days in the historical one. The groups differ in how often they take the GRA-linked path (82.5%, 80.1%, and 44.3%) and in how long each route takes: GRA-linked means are 21.65, 19.67, and 23.79 days, and NVD-first means are 23.35, 18.59, and 44.34 days. In what follows, we further analyze the mix of route choices and within-route timings.

Cohorts, paths, and outcome

The AI-associated candidates correspond to 469 validated advisories identified by our mining pipeline. For the route analysis we retain the 468 on the GRA-linked (386) and NVD-first (82) paths and set the singleton residual advisory aside, as in Section 4.3. Both comparison groups are drawn from the advisories table of the same snapshot, which holds every reviewed GHSA. We call an advisory a non-candidate when it is not among the 469 valid candidates; the 29 rejected linkages therefore belong to the references. Absence from the candidate list is not proof that no AI tool was involved. Like the whole analysis, the comparison is restricted to reviewed advisories.

The contemporaneous reference CC contains the 12,951 non-candidates (9,578 GRA-linked, 3,373 NVD-first) published in the 19 months that contain at least one candidate. Each month is reweighted to the share it has among the 468 candidates, split evenly across that month’s advisories, so that CC has the candidates’ calendar mix. The historical reference HH contains the 10,403 non-candidates (4,612 GRA-linked, 5,791 NVD-first) published from July 2022, after the NVD backfill transition, to August 2025, the period covered by the earlier GHSA study [26]; it is unweighted. Both references exclude advisories without a source time or with a source time later than the review (4 and 8 in CC, 6 and 66 in HH), and we write AA for the AI-associated candidates.

A GRA-linked advisory contains a repository-advisory URL. An NVD-first advisory lacks that link and contains an NVD publication timestamp. The two-route restriction places the singleton candidate residual category aside and avoids assigning a common mechanism to heterogeneous remaining sources. Figure 7 summarizes the observable paths and group sizes.

Two disclosure paths retained in the comparative analysis

Figure 7. The two disclosure paths retained in the comparative analysis. AA denotes AI-associated candidates, CC the contemporaneous reference, and HH the historical reference; L~\widetilde{L} denotes the route-specific (weighted) median review lag in days, measured from the source publication (repository advisory or NVD record) to GitHub’s review.

We measure the outcome from the source publication:

L=treview−tsource,(1)L = t^{\mathrm{review}} - t^{\mathrm{source}}, \tag*{(1)}

where treviewt^{\mathrm{review}} is GitHub’s review time. Here, tsourcet^{\mathrm{source}} denotes the start of the review-lag clock: the publication time of the GitHub Repository Advisory for a GRA-linked advisory, or the publication date recorded by the NVD for an NVD-first advisory. Thus, the clock begins when the advisory becomes public through the source associated with its route into the GHSA review pipeline, corresponding to the source-to-review clock of Segal et al. [26].

Potential changeEffect on review demandPossible consequence
Lower cost of preparing a changeMore proposed changes, including redundant or unsolicited onesMore work enters the review queue
Faster patch productionSubmissions arrive in a shorter periodTemporary congestion if review capacity does not grow at the same rate
Limited project context or unclear rationaleMore effort to assess relevance, design fit, and correctnessLonger triage or review for some changes
Triage policies and contribution controlsSome proposals are filtered, redirected, or delayed before formal reviewReview demand may be contained, depending on policy effectiveness

Table 4. A plausible pathway from coding-agent use to review demand. The mechanisms are interpretive and were not directly measured in this study.

The impact of AI-associated candidates

The AI-associated candidates have longer median review lags than the same-month reference within both disclosure routes: 7.4 versus 6.0 days on the GRA-linked route, and 12.0 versus 7.4 days on the NVD-first route. Their mean lag is 2.49 days higher, although the confidence interval for this difference spans zero, and their route shares are nearly identical to those of the same-month reference. The candidates were reviewed sooner than their same-month counterparts in 45.8% of GRA-linked comparisons and 42.5% of NVD-first comparisons. Thus, the medians are higher for the candidates in both routes, while their review-time distributions substantially overlap with those of the same-month reference.

Review lags also increased during 2026 for GRA-linked advisories generally: the median interval from repository-advisory publication to the appearance of the reviewed GHSA rose from 0.4 days in January to 5.7 days in May and 41.4 days in September. The candidates exhibit the same broad temporal pattern.

Table 5 shows the median review lag for contemporaneous and historical advisories. One possible mechanism that explains the recent increase in median review lag compared to the historical reference is that coding agents lower the cost of preparing and submitting changes, allowing the number of proposed changes to grow faster than maintainers’ capacity to assess them. Faster production at the submission stage need not mean faster review: maintainers still need to establish whether a change is relevant, understand its context, check its correctness and tests, and decide whether it fits the project’s design. If submissions arrive faster than they can be triaged and reviewed, the queue of reviews may grow and typical waiting times may increase.

RoutennShareMeanP25MedianP75P90≤1\le1 day
AI-associated candidates (A)
GRA-linked38682.5%21.651.327.3539.3364.1217.4%
NVD-first8217.5%23.353.6211.9841.2459.3812.2%
Contemporaneous reference (C)
GRA-linked9,57880.1%19.671.256.0034.0056.0121.2%
NVD-first3,37319.9%18.591.237.4231.2350.0719.8%
Historical reference (H)
GRA-linked4,61244.3%23.790.180.451.183.2269.9%
NVD-first5,79155.7%44.340.331.105.1055.2847.5%

Table 5. Review lag (source publication to review) by group and disclosure route. The residual route is excluded. Counts are unweighted; shares, means, and quantiles use the group weights (the contemporaneous reference is month-weighted). Timing columns are in days.

Route-specific numerical comparison

Table 5 reports the route-specific distributions of the three groups. The GRA-linked share of the candidates (82.5%) is close to that of CC (80.1%) and 38.2 percentage points above that of HH (44.3%). On both routes the candidates’ lags are long, and the contemporaneous reference is the closest comparison.

The group means are reconstructed, up to rounding, by their route mixtures:

μA=0.825(21.65)+0.175(23.35)=21.95 days,(2)\mu_A = 0.825(21.65) + 0.175(23.35) = 21.95\ \text{days}, \tag*{(2)}
μC=0.801(19.67)+0.199(18.59)=19.45 days,(3)\mu_C = 0.801(19.67) + 0.199(18.59) = 19.45\ \text{days}, \tag*{(3)}
μH=0.443(23.79)+0.557(44.34)=35.23 days.(4)\mu_H = 0.443(23.79) + 0.557(44.34) = 35.23\ \text{days}. \tag*{(4)}

The historical reference differs from the others in both location and tail. Its GRA-linked and NVD-first medians (0.45 and 1.10 days) are a small fraction of the candidates’ (7.35 and 11.98 days), and 69.9% and 47.5% of its advisories are reviewed within one day, versus 17.4% and 12.2% of the candidates. Yet its NVD-first mean (44.34 days) is almost twice the candidates’ (23.35 days), because of an extreme tail: the longest lag on either route is 91.9 days among the candidates, 1,852.1 days in CC, and 2,319.1 days in HH. As reported in § 4.3, a candidate advisory is reviewed sooner than a contemporaneous one in 45.8% of GRA-linked and 42.5% of NVD-first pairs, and sooner than a historical one in 16.3% and 25.5%. Figure 5 shows both the central shift and the tails.

Exact mean decomposition

Next, we decompose the difference in mean review lag between advisories using AI coding agents and each reference group into two components: the contribution associated with differences in the proportions following the GRA-linked and NVD-first routes, and the contribution associated with differences in mean review lag within each route.

Let ZZ denote the set of advisories of interest, where Z∈{A,B}Z \in\{A, B\}. Here, AA refers to the AI-associate candidates and BB refers to the baseline reference group, B∈{C,H}B \in\{C, H\}, where CC and HH refer to the contemporaneous and historical advisories, respectively. Let R∈R={G,N}R \in\mathcal{R} = \{G, N\} denote the GRA-linked and NVD-first routes. With the group weights of Section 5.1, define

qz,r=Pr⁡(R=r∣Z=z),μz,r=E[L∣Z=z,R=r].(5)q_{z,r} = \operatorname{Pr}(R = r \mid Z = z), \qquad\mu_{z,r} = \mathbb{E}[L \mid Z = z, R = r]. \tag*{(5)}

The group mean is the finite mixture

μz=E[L∣Z=z]=∑r∈Rqz,rμz,r.(6)\mu_z = \mathbb{E}[L \mid Z = z] = \sum_{r \in\mathcal{R}} q_{z,r}\mu_{z,r}. \tag*{(6)}

The mean review lag difference Δ=μA−μB\Delta= \mu_A - \mu_B can be decomposed as

Δ=∑r∈R(qA,r−qB,r)μB,r⏟Δroute+∑r∈RqA,r(μA,r−μB,r)⏟Δwithin.(7)\Delta= \underbrace{\sum_{r \in\mathcal{R}} (q_{A,r} - q_{B,r})\mu_{B,r}}_{\Delta_{\mathrm{route}}} + \underbrace{\sum_{r \in\mathcal{R}} q_{A,r}(\mu_{A,r} - \mu_{B,r})}_{\Delta_{\mathrm{within}}}. \tag*{(7)}

The first term changes route composition while retaining the reference’s route-specific timings; the second changes route-specific timings while retaining the candidate route mix, and splits into one contribution per route.

Indeed, to verify the above equality note that

μA−μB=∑rqA,rμA,r−∑rqB,rμB,r=(∑rqA,rμB,r−∑rqB,rμB,r)+(∑rqA,rμA,r−∑rqA,rμB,r),\begin{aligned} \mu_A - \mu_B = \sum_r q_{A,r}\mu_{A,r} - \sum_r q_{B,r}\mu_{B,r} \\ &= \left(\sum_r q_{A,r}\mu_{B,r} - \sum_r q_{B,r}\mu_{B,r}\right) + \left(\sum_r q_{A,r}\mu_{A,r} - \sum_r q_{A,r}\mu_{B,r}\right), \end{aligned}

which yields Eq. (7).

Table 6 and Figure 8 report the estimates against both references. Confidence intervals come from 1,000 bootstrap replicates that resample repositories independently within each group, preserving within-project dependence; they use the resampling of § 4.3, including month weights recomputed from the resampled candidates for CC. Each comparison has its own resamples, so the interval of μA\mu_A differs slightly between the two columns.

QuantityContemporaneous (B=CB = C) EstimateContemporaneous (B=CB = C) 95% intervalHistorical (B=HB = H) EstimateHistorical (B=HB = H) 95% interval
Reference mean, μB\mu_B19.45[16.38, 22.60]35.23[26.53, 46.76]
AI-associated mean, μA\mu_A21.95[16.78, 27.19]21.95[16.23, 27.33]
Marginal difference, Δ\Delta+2.49[−1.72-1.72, 7.20]−13.28-13.28[−25.36-25.36, −2.90-2.90]
Route composition, Δroute\Delta_{\mathrm{route}}+0.03[−0.19-0.19, 0.31]−7.84-7.84[−16.57-16.57, −1.33-1.33]
Within-route timing, Δwithin\Delta_{\mathrm{within}}+2.47[−1.75-1.75, 7.15]−5.44-5.44[−14.24-14.24, 2.92]
GRA-linked contribution+1.63[−2.63-2.63, 6.08]−1.77-1.77[−10.17-10.17, 5.98]
NVD-first contribution+0.83[−0.33-0.33, 2.13]−3.68-3.68[−7.81-7.81, −0.91-0.91]

Table 6. Route–timing decomposition of the review lag in days against each reference group. Brackets show 95% repository-cluster bootstrap intervals.

Numerical route–timing decomposition against contemporaneous and historical references

Figure 8. Numerical route–timing decomposition against the contemporaneous (left) and historical (right) references. The intermediate mean applies the AI-associated route shares to the reference’s route-specific means. Brackets show 95% repository-cluster bootstrap intervals.

Against the contemporaneous reference the candidates’ mean is 2.49 days higher, with an interval that spans zero. Their route mix is essentially that of CC (Δroute=+0.03\Delta_{\mathrm{route}} = +0.03), so the whole difference is within-route timing, mostly on the GRA-linked route, and it is not distinguishable from zero. Against the historical reference the candidates are 13.28 days faster on the marginal mean, and 7.84 days (59.0%) of that difference is route composition: replacing the historical shares (44.3% GRA, 55.7% NVD-first) with the candidate shares (82.5% GRA, 17.5% NVD-first) while holding the historical route-specific means fixed lowers the expected lag from 35.23 to 27.39 days. The within-route component is −5.44-5.44 days; its confidence interval includes zero, and its sign depends on the historical tail, as further detailed next.

Distributional interpretation

Next, we extend our analysis beyond mean and median review lags, to account for the distribution of the review lag. Let Sz,r(t)=Pr⁡(L>t∣Z=z,R=r)S_{z,r}(t) = \Pr(L > t \mid Z = z, R = r) be the route-specific survival function. Then

Sz(t)=∑r∈Rqz,rSz,r(t).(8)S_z(t) = \sum_{r \in\mathcal{R}} q_{z,r}S_{z,r}(t). \tag*{(8)}

A route-standardized candidate curve under the reference’s route mix is

SAstd(t)=∑r∈RqB,rSA,r(t).(9)S_A^{\mathrm{std}}(t) = \sum_{r \in\mathcal{R}} q_{B,r}S_{A,r}(t). \tag*{(9)}

Comparing SAstdS_A^{\mathrm{std}} with SBS_B removes the marginal route-share difference while retaining within-route distributional differences. Under the historical route mix, the standardized candidate curve has 14.5% of advisories reviewed within one day, against 57.5% for HH, and 37.2% with a lag above 30 days, against 7.7%. Under the contemporaneous mix the gaps are smaller but point the same way: 16.3% versus 20.9% within one day, and 34.8% versus 26.8% above 30 days. Thus, within routes, the bulk of the AI candidates’ distribution is slower than both contemporaneous and historical references.

The mean points the other way against HH because it is dominated by lags that the AI-associated candidates cannot match: the longest AI-associated candidate review lag is 91.9 days, while HH reaches 2,319.1. As a sensitivity check, we cap every lag at 91.9 days, the longest AI-associated candidate lag; the cap is used to expose the role of the tail. Against HH, the capped historical mean falls from 35.23 to 8.25 days, the route component remains negative (−2.85 days, interval [−4.40, −1.69]), and the within-route component changes sign to +16.55 days ([10.84, 22.25]), driven mainly by the GRA-linked route (+14.48 days, [8.86, 20.42]); the capped marginal difference is +13.69 days ([7.97, 19.27]). Against CC the capped within-route component is +4.03 days ([−0.08, 8.29]) and the capped marginal difference is +4.03 days ([−0.08, 8.23]). The negative within-route mean against HH is therefore produced by the extreme historical tail rather than by faster typical review of the AI-associated candidates.

Interpretation, calendar trend, and scope

Relative to the historical post-backfill data, AI-associated candidates have a lower marginal mean review lag primarily because they are much more likely to travel through the GRA-linked path. Relative to advisories published in the same months, they have no clearly different mean, and their route mix is the same. Neither comparison supports the claim that AI accelerates review within a fixed route: within routes the median review lag is larger than in both references.

The historical contrast is confounded with calendar time, because review lags lengthened substantially across the whole platform. Among all valid non-candidate GRA-linked advisories, the median lag is 0.4 days for advisories published in January 2026, 5.7 days in May, 29.3 days in July, and 41.4 days in September; the candidates’ medians for the same months are 0.5, 4.5, 34.6, and 32.4 days. Because 431 of the 469 valid candidates were published in 2026, month standardization is essential, which is why the contemporaneous reference is the primary comparison.

As a second sensitivity analysis we repeat the decomposition on the same groups with the interval from GHSA publication to review. GRA-linked lags are then zero, and only the NVD-first terms remain. The marginal difference is +0.47 days against C ([-1.69, 2.94]), with route composition −0.43 ([-1.84, 1.23]) and within-route timing +0.90 ([-0.27, 2.16]); against H it is −16.94 days ([-29.06, −8.28]), of which 85.0% is route composition (−14.40 days, [-23.49, −8.06]) and the within-route component is −2.53 days ([-6.55, 0.24]). The qualitative conclusion is unchanged, but the GHSA clock understates the lag of the GRA-linked route and, with it, the weight of the calendar trend.

Discussion

Evidence and scale of AI coding agents in GHSAs

Our results evidence the existence and scale of AI coding agents in GHSAs: hundreds of reviewed advisories now reference commits whose public Git metadata explicitly attributes work to coding agents. The evidence spans general-purpose agents and security-specific automation such as Copilot Autofix.

The evidence does not establish how much code an agent produced, whether a human substantially rewrote it, or whether the final change was correct because of the agent. Indeed, co-author trailers are simple declarations. Exact bot identities are stronger operational evidence but still do not reveal the division of labor. Instruction files and configuration directories are weaker still: they show that maintainers prepared a repository for agent interaction, not that a particular security fix was agent-authored.

The timing result is equally specific. AI-associated candidates have a lower marginal mean lag than the historical cohort (13.28 days), and 59% of that difference is their much greater use of the GRA path. The rest, a negative within-route component, comes from the historical extreme tail and reverses sign when lags are capped. Against advisories published in the same months, the candidates have the same route mix and no clearly different mean (+2.49 days, interval spanning zero). Within each route, advisories using coding agents have a higher median review lag than advisories in either reference group. On the GRA-linked route, the median is 7.4 days, compared with 6.0 days in the same-month reference and 0.4 days in the historical reference; on the NVD-first route, it is 12.0 days, compared with 7.4 and 1.1 days, respectively. These data therefore do not support claims that AI-assisted review is faster.

Throughput constraint. The more striking pattern is platform-wide. The median lag between the publication of a repository advisory and its reviewed GHSA rose from 0.4 days for advisories published in January 2026 to 5.7 days in May and 41.4 days in September, and the AI-associated candidates track it. Public timestamps alone cannot identify the cause, but GitHub has reported that inbound advisory volume rose sharply in 2026 (1,560 reviewed advisories in May alone, over five times its typical monthly output) and that review times for new advisories lengthened as a result (Git 2026). The direction of our measured trend is consistent with this account of a throughput constraint.

Implications for maintainers and platforms

Security advisories would benefit from structured provenance. Today, attribution is scattered across free-form trailers, bot identities, pull request text, and repository files. A structured field distinguishing generation, review, testing, and documentation roles would reduce ambiguity and enable more reliable audits. Such provenance should identify the human maintainer who accepted the change and should not imply that an automated tool bears responsibility for disclosure decisions.

Maintainers can improve traceability by preserving agent attribution through squash merges, requiring security tests for agent-assisted fixes, and recording whether the agent generated the patch, suggested it, reviewed it, or merely edited supporting text. For high-impact vulnerabilities, repository policy could require a human-authored rationale explaining the vulnerable behavior, the security invariant restored by the patch, and the regression tests that exercise it.

For GitHub and other advisory platforms, the concentration of candidate observations in GRAs, which mirrors that of all advisories published in the same months, suggests that repository-native disclosure is the natural place to capture provenance. Optional structured metadata could support research and governance without exposing private prompts or confidential draft advisories.

Threats to Validity

Construct validity. Git author, committer, and co-author fields are unauthenticated. An exact trailer can be spoofed, omitted, or added by a human after the agent’s work. Conversely, an agent contribution may leave no detectable signal. Our rules favor precision, so the corpus is a lower-bound view of visible attribution, not an estimate of all agent participation. The identified Devin human-name collision demonstrates that even seemingly distinctive tool names require validation. Artifact-only candidates are reported separately because they do not support commit-level authorship.

Link validity. An advisory reference is not guaranteed to identify the unique patch commit. References may point to a release, test, merge, backport, documentation change, or related discussion. One commit can support multiple advisories, and one advisory can cite multiple commits. We therefore report all counting units and avoid equating an advisory–commit link with vulnerability introduction. Diff-level classification is needed for a stronger remediation-role taxonomy.

Internal validity. The analysis is observational. Repository maturity, activity, ecosystem, security policy, project popularity, and calendar time may jointly influence agent use, GRA adoption, and review latency. In particular, 431 of 469 valid candidates were published in 2026, whereas the historical comparison ends in August 2025, and review lags lengthened markedly in between. The decomposition reveals whether a marginal difference is associated with route shares or within-route means, and the same-month reference removes the calendar-month confound, but neither removes project-composition confounding or identifies causal effects.

Timing validity. The review lag starts at the source publication, dated by the repository-advisory publication or the NVD publication; private coordination before that moment is not observable. The interval from GHSA publication to review gives the same qualitative conclusions but is exactly zero for GRA-linked advisories, which hides the delay we measure. The analysis covers only advisories that GitHub had reviewed at collection, so long lags are less observable for recent publications; the same-month reference shares this truncation by construction, but the historical one does not.

External validity. The study includes public GitHub repositories linked to reviewed GHSAs. Private repositories, unreviewed advisories, proprietary incident systems, and agents that do not preserve public attribution are outside scope. Tool-family frequencies reflect both use and observability; they should not be generalized to overall market share or relative security quality.

Reliability. Scans are pinned to immutable targets and findings use stable identities. Nevertheless, the ecosystem evolves rapidly: new agent accounts, trailer conventions, and configuration paths will require rule updates. We will publish the rule catalog, snapshot date, analysis code, and candidate table so that later replications can distinguish true temporal change from detector evolution.

Conclusion

We presented a security-advisory-centered study of coding agents in real software repositories. By joining high-precision local-Git findings to commits referenced by reviewed GHSAs, we obtained 469 valid candidate advisories across 189 repositories after removing 29 advisories affected by human-name collisions. Direct attribution appears in 448 advisories, while 21 are supported by repository instruction or configuration artifacts. The candidates cover a broad range of severity and weakness classes, and most are connected to the fast review, repository-native GRA path, as are other advisories published in the same months.

The main methodological lesson is that attribution must be layered. A bot identity, an explicit co-author trailer, and an AGENTS.md file are not interchangeable evidence. The main analytical lesson is that GHSA route composition between the NVD and GRA paths must be separated from within-route timing before differences in review latency are interpreted. In addition, we observed interesting synchronization effects: dating a GRA-linked advisory from its GHSA publication yields a negligible review lag, as publication and review typically occur concomitantly for GRA-first advisories, whereas dating it from the repository advisory exposes a delay that grew across the platform during 2026. Measured from source publication, the AI-associated candidates’ mean review lag is 13.28 days lower than the historical data and 2.49 days higher than that of advisories published in the same months. Of the historical difference, 7.84 days come from their 82.5% GRA share versus 44.3% historically, and the within-route component depends on the historical extreme tail; against same-month advisories the route mix is the same and the difference is not clear. Within each route, the AI-associated candidates are not faster, suggesting that the increase in the reported inbound advisory volume and throughput constraints may play a role in the review delays.

Future work should connect repository-level evidence to the full remediation and review process. Linking agent-attributed commits to pull requests could allow researchers to measure submission volume, time to first review, review duration, and time to merge, while distinguishing direct agent contributions from repository-level adoption signals. Longitudinal data could then test whether changes in agent use are associated with changes in review workload and delay, accounting for route, publication month, and advisory characteristics. Public traces now make coding-agent involvement in security remediation measurable; our work is a first step towards a research agenda that aims to understand what happens throughout the review process and to establish whether involvement of AI coding agents leads to faster or more effective remediation.

Replication and Data Availability

The replication package includes a versioned catalog of detectors of AI coding agents, schema, notebooks, deduplicated candidate table, validation decisions, and scripts that produce the tables and figures. It is archived on Zenodo at https://doi.org/10.5281/zenodo.23378198.

Detection-rule catalog

Table 7 lists the values recognized by each detection rule summarized in Table 1, together with the implementation identifier.

RuleRecognized values
Message signal (commit_ai_evidence.message-signal)Line-anchored Co-authored-by: Name <email> trailers whose name is an allowlisted bot handle, contains one of the whole tokens copilot, claude, codex, chatgpt, gemini, cursor, devin, aider, coderabbit, sweep, codeium, sourcery, or codeflash, or whose address is a canonical Copilot, Anthropic, or OpenAI address. Explicit generation phrases (e.g., “generated by GitHub Copilot,” “Claude Code created,” “AI-generated”). Workflow forms (a standalone “Apply Copilot suggestion” line; “Copilot Autofix”).
Known bot identity (commit_bot_author.known-ai-account)Author or committer name equal, case-insensitively, to github-copilot[bot], copilot-swe-agent[bot], devin-ai-integration[bot], sweep-ai[bot], coderabbitai[bot], sourcery-ai[bot], or codeflash-ai[bot]. E-mail addresses are not used to infer identity.
Instruction file (instruction_file.universal)Addition of, or rename into, one of the root files AGENTS.md, CLAUDE.md, CONTEXT.md, GEMINI.md, CODEX.md, COPILOT.md, CURSOR.md, CLINE.md, QWEN.md, or JULES.md (exact, case-sensitive). For a rename, the previous path is retained.
Agent configuration (agent_config_file.universal)Addition of, or rename into, one of the root files .cursorrules, .clinerules, .aider.conf.yml, .aiderignore, .aider.model.settings.yml, opencode.json, or opencode.jsonc; or addition of any file below one of the top-level directories .claude, .cursor, .gemini, .codex, .windsurf, .amazonq, .continue, .opencode, or .qwen, recorded at the directory level.

Table 7. Recognized values per detection rule.

Examples of project contribution policies discussed in response to AI-generated submissions

Public discussion of AI-assisted contributions includes projects that have restricted outside pull requests or introduced additional conditions for accepting them (see Table 8). These examples provide context for the possibility that lower-cost change production can increase the demand placed on human review (see § 6.1).

ProjectReported policyInterpretation
LadybirdThe contribution guide states that pull requests from non-maintainers are not accepted.Restriction on outside pull requests; the policy itself does not establish that AI was its sole cause.
tldrawPublic reporting describes automatic closure of external pull requests and redirection of proposals to issues or discussions.Restriction on the external PR channel.
FluePublic reporting describes automatic closure of external pull requests, with proposals redirected to issues or discussions.Restriction on the external PR channel.
GhosttyThe contribution guide asks that pull requests be associated with a previously accepted issue, and its AI policy sets conditions for AI-assisted contributions.Conditional contribution policy, rather than a blanket ban on pull requests.

Table 8. Illustrative project policies concerning outside or AI-assisted contributions. Policies may change over time.

Reported policies differ. tldraw and Flue have been described as automatically closing external pull requests and redirecting proposals to issues or discussions. Ladybird announced that it would stop accepting public pull requests. Ghostty, by contrast, has been reported to permit AI-assisted pull requests under conditions such as prior agreement on an issue, while rejecting unsolicited submissions. Some restrict all external pull requests, whereas others restrict only certain AI-assisted or unsolicited contributions. Project policies also evolve, so the examples are a snapshot of public reporting.

Declarations

Funding This work was partially supported by INES.IA (National Institute of Science and Technology for Software Engineering Based on and for Artificial Intelligence), www.ines.org.br, CNPq grant 408817/2024-0. This work was also partially supported by CAPES and FAPERJ under grants E-26/204.268/2024 and E-26/260.168/2026, CNPq grants 444956/2024-7, 424622/2021-1 and 315106/2023-9, as well as Finep PlatCiber.

Ethical Approval The study analyzes public software repositories and public security-advisory metadata and does not involve human-subject interaction. Evidence is reported in aggregate except for public repository identifiers already contained in the replication dataset.

Informed Consent Not applicable.

Conflict of Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Clinical Trial Number Not applicable.

Author Contributions. Claudio Segal and Paulo Segal contributed data curation, software, and investigation. Hudson Silva Borges contributed methodology, data collection, validation and interpretation. Daniel S. Menasché contributed conceptualization, analytical modeling, supervision, and writing. Paulo Silveira, Eduardo de Almeida, Felipe Paixão, Joanna C.S. Santos and Raphael Machado contributed writing and mentoring. Anton Kocheturov and Gaurav Srivastava contributed industrial insights. All authors reviewed and approved the manuscript.

References

  1. [2]Alexopoulos N, Meneely A, Arnouts D, Mühlhäuser M (2021) Who are vulnerability reporters? A large-scale empirical study on FLOSS. In: Proceedings of the 15th ACM/IEEE international symposium on empirical software engineering and measurement (ESEM), pp 1–12
  2. [3]Allodi L, Massacci F (2017) Security events and vulnerability data for cybersecurity risk estimation. Risk Analysis 37(8):1606–1627DOI
  3. [4]Arora A, Krishnan R, Telang R, Yang Y (2010) An empirical analysis of software vendors’ patch release behavior: impact of vulnerability disclosure. Information Systems Research 21(1):115–132DOI
  4. [5]Ayala J, Tung YJ, Garcia J (2024) Poster: A glimpse of vulnerability disclosure behaviors and practices using GitHub projects. In: Proceedings of the 45th IEEE Symposium on Security and Privacy (S&P)
  5. [6]Ayala J, Ngo S, Garcia J (2025a) A deep dive into how open-source project maintainers review and resolve bug bounty reports. In: 2025 IEEE Symposium on Security and Privacy (SP), IEEE, pp 522–538DOI
  6. [7]Ayala J, Tung YJ, Garcia J (2025b) A Mixed-Methods Study of Open-Source Software Maintainers On Vulnerability Management and Platform Security Features. In: 34th USENIX Security Symposium (USENIX Security 25), pp 2105–2124arxiv.org/abs/2409.07669
  7. [8]Ayala J, Tung YJ, Garcia J (2025c) Investigating vulnerability disclosures in open-source software using bug bounty reports and security advisories. arXiv preprint arXiv:250117748arxiv.org/abs/2501.17748
  8. [9]Bozorgi M, Saul LK, Savage S, Voelker GM (2010) Beyond heuristics: learning to classify vulnerabilities and predict exploits. In: Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining, pp 105–114
  9. [10]Decan A, Mens T, Constantinou E (2018) On the impact of security vulnerabilities in the npm package dependency network. In: Proceedings of the 15th international conference on mining software repositories, pp 181–191DOI
  10. [11]Fu Y, Liang P, Tahir A, Li Z, Shahin M, Yu J, Chen J (2025) Security Weaknesses of Copilot-Generated Code in GitHub Projects: An Empirical Study. URL https://arxiv.org/abs/2310.02059, 2310.02059
  11. [12]GitHub (2025) Github advisory database: Security vulnerability database inclusive of cves and github-originated security advisories. URL https://github.com/advisories, see also https://github.com/github/advisory-database and https://docs.github.com/en/code-security/security-advisories/working-with-repository-security-advisories/about-repository-security-advisories. Accessed: 2025-08-18
  12. [13]GitHub Blog (2022) All historical nvd advisories are now listed on github. URL https://github.blog/changelog/2022-06-08-all-historical-nvd-advisories-are-now-listed-on-github/, accessed: 2025-09-12
  13. [15]GitHub Docs (2025b) About the github advisory database. URL https://docs.github.com/en/code-security/security-advisories/working-with-global-security-advisories-from-the-github-advisory-database/about-the-github-advisory-database, accessed: 2025-10-16
  14. [16]He R, He H, Zhang Y, Zhou M (2023) Automating dependency updates in practice: An exploratory study on GitHub Dependabot. IEEE Transactions on Software Engineering 49(8):4004–4022arxiv.org/abs/2206.07230
  15. [18]Kinsman T, Wessel M, Gerosa MA, Treude C (2021) How do software developers use github actions to automate their workflows? In: 2021 IEEE/ACM 18th International Conference on Mining Software Repositories (MSR), IEEE, pp 420–431DOI
  16. [19]Li F, Paxson V (2017) A large-scale empirical study of security patches. In: Proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, pp 2201–2215
  17. [20]Li H, Zhang H, Hassan AE (2025) The rise of AI teammates in software engineering (SE) 3.0: Characterizing agentic coding in GitHub pull requests. arXiv preprint arXiv:250715003 DOI 10.48550/arXiv.2507.15003
  18. [21]Liu S, Zhou J, Hu X, Gogo FR, Xia X, Yang X (2025) An empirical study on vulnerability disclosure management of open source software systems. ACM Transactions on Software Engineering and MethodologyDOI
  19. [22]Mirhosseini S, Parnin C (2017) Can automated pull requests encourage software developers to upgrade out-of-date dependencies? In: 2017 32nd IEEE/ACM international conference on automated software engineering (ASE), IEEE, pp 84–94DOI
  20. [23]Pearce H, Ahmad B, Tan B, Dolan-Gavitt B, Karri R (2025) Asleep at the keyboard? assessing the security of github copilot’s code contributions. Communications of the ACM 68(2):96–105
  21. [24]Ponta SE, Plate H, Sabetta A (2020) Detection, assessment and mitigation of vulnerabilities in open source dependencies. Empirical Software Engineering 25(5):3175–3215DOI
  22. [25]Rooijendijk P, Treude C, Wessel M (2026) Who said CVE? how vulnerability identifiers are mentioned by humans, bots, and agents in pull requests. In: Proceedings of the 23rd International Conference on Mining Software Repositories, ACM, New York, NY, USA, DOI 10.1145/3793302.3793616, arXiv:2601.19636DOI
  23. [26]Segal C, Segal P, Banjar CE, Paixão FdS, Borges HS, Silveira P, Almeida ESd, Santos JCS, Kocheturov A, Srivastava G, Menaschè DS (2026) Characterizing and modeling the GitHub security advisories review pipeline. In: Proceedings of the 23rd International Conference on Mining Software Repositories, pp 311–322, DOI 10.1145/3793302.3793360DOI
  24. [27]Siddiq ML, Zhao X, Lopes VC, Casey B, Santos JCS (2026) Security in the Age of AI Teammates: An Empirical Study of Agentic Pull Requests on GitHub. URL https://arxiv.org/abs/2601.00477, 2601.00477
  25. [28]Wang B, Li X, de Aguilar LP, Menasche DS, Shafiq Z (2017) Characterizing and modeling patching practices of industrial control systems. Proceedings of the ACM on Measurement and Analysis of Computing Systems 1(1):1–23
  26. [29]Zimmermann M, Staicu CA, Tenny C, Pradel M (2019) Small world with high risks: A study of security threats in the npm ecosystem. In: 28th USENIX Security symposium (USENIX security 19), pp 995–1010arxiv.org/abs/1902.09217

Paper details

vlnc:W_ftu8jr99
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents