WorksComputer Science
We Have All the Cards: Securing Agent Cards at the Operational Trust Boundary of the Agent2Agent Protocol
Abstract
AI agent interoperability protocols enable orchestrators to discover and invoke external agents using machine-readable metadata. In the Agent2Agent (A2A) protocol, Agent Cards contain attacker-controllable natural-language fields that may influence agent selection before a provider is invoked, creating a pre-invocation trust problem. We characterize this discovery-time attack surface through a taxonomy of ten Agent Card manipulation attacks and present SECCARD, a layered validation framework combining protocol checks, deterministic analysis, and LLM-based semantic reasoning. We measure the emerging public A2A ecosystem by probing 36.1 million Internet domains and collecting 310 deployed Agent Cards. Two annotators independently label the corpus with 96.8% agreement. We then construct a 660-card real-card perturbation benchmark with matched benign controls and a 260-card synthetic adversarial benchmark spanning all ten attack categories. On controlled attacks injected into real Agent Cards, the strongest semantic judge achieves 97.0% security recall at a 1.7% false-positive rate. Across nine LLM judges on the synthetic benchmark, security recall ranges from 25.8% to 92.3%, demonstrating substantial model-dependent variation. These results show that SECCARD effectively identifies Agent Card manipulations spanning all ten discovery-time attack classes while preserving benign Agent Cards, but also reveal that its effectiveness depends critically on the choice of the underlying LLM judge.
Introduction
Advances in Large Language Models (LLMs) have accelerated the development of autonomous LLM-based AI agents capable of planning and multi-step reasoning. As these systems grow in complexity, organizations are adopting orchestrator agents that coordinate multiple specialized agents to collaboratively solve complex tasks [1], [2].
This architectural shift is rapidly moving into production. In a 2026 survey, 57% of professionals reported having AI agents already deployed in production [3].
This has motivated the development of standardized interoperability protocols that enable agent discovery, capability advertisement, and communication across agents. Among these, the Agent2Agent (A2A) protocol, released in April 2025 by Google, has emerged as one of the first open standards designed for agent interoperability [4].
A fundamental challenge faced by agent interoperability protocols is agent discovery: before an orchestrator can delegate a task, it must discover candidate agents and determine which one is most suitable, i.e. possesses the required skill(s) to complete the task. Unlike traditional service-oriented architectures, where endpoints are configured manually, agent orchestrators require machine-readable metadata that describe an agent’s identity, capabilities, skills, authentication requirements, and communication endpoints.
To solve this challenge, the A2A protocol allows organizations to publish JavaScript Object Notation (JSON)-based agent cards to advertise their capabilities to other agents [5], [6]. An orchestrating agent that wishes to delegate a task fetches the publicly available card file, parses its fields and decides whether to invoke the advertised service, all without human review.
The security implication is immediate: agent card content is untrusted text that an LLM consumes as input. An attacker who controls a domain or compromises a legitimate one can publish a card whose text fields carry prompt injection payloads targeted at the ingesting orchestrator. Unlike web content that humans evaluate before acting on, agent cards are designed to drive autonomous LLM-based agent decisions that may trigger real-world actions: transferring funds, exfiltrating conversation context, or silently escalating invocation privileges. The A2A specification provides no normative guidance on how to evaluate card trustworthiness beyond schema conformance.
Research gap. Indirect prompt injection through externally controlled content is a well-established threat to LLM-integrated systems [7]–[10]. Recent work has extended this concern to A2A systems, identifying Agent Card spoofing, capability cloaking, and card poisoning as potential attacks on agent discovery.
However, existing work characterizes only a limited set of discovery-time Agent Card manipulations and primarily evaluates attacks in controlled settings. We still lack a fine-grained understanding of the ways adversarial Agent Cards can manipulate discovery decisions, empirical evidence about such manipulations in deployed Agent Cards, and a dedicated pre-invocation validation approach evaluated against both adversarial and benign cards. This gap is operationally important because Agent Cards are not merely external content retrieved by an LLM; they are metadata explicitly designed to influence which external agents an orchestrator selects and invokes. Consequently, adversarial content can affect the discovery decision before the advertised provider enters the workflow.
Our approach. We study Agent Card security at the discovery-to-invocation trust boundary, where attacker-controlled metadata can influence an orchestrator before the advertised provider is invoked. We develop a taxonomy of ten discovery-time Agent Card manipulation attacks, characterizing how adversarial cards can manipulate agent selection, misrepresent capabilities, and evade validation.
Building on this taxonomy, we introduce SECCARD, a layered pre-invocation validation framework. Layer 1 validates protocol conformance against the A2A specification. Layer 2A applies deterministic NLP analysis to identify lexical and structural indicators of manipulation, while Layer 2B employs an LLM-as-a-judge semantic classifier [11] to identify attacks that require reasoning about meaning, intent, and consistency.
We evaluate SECCARD using three complementary datasets: 310 real-world Agent Cards collected through an Internet-scale measurement, a 660-card real-card perturbation benchmark with matched benign controls, and a 260-card synthetic adversarial benchmark spanning all ten attack categories. On controlled attacks injected into real Agent Cards, the strongest semantic judge achieves 97.0% security recall at a 1.7% false-positive rate. Across nine LLM judges on the synthetic benchmark, security recall ranges from 25.8% to 92.3%. Together, these experiments evaluate Agent Card validation across naturally occurring deployments, controlled attacks on real cards, and broader adversarial variation.
Contributions. This paper makes the following contributions:
Internet-scale measurement of deployed Agent Cards. We probe 36.1 million Internet domains and collect 310 publicly deployed Agent Cards, providing an empirical view of the emerging A2A ecosystem. Two annotators independently establish security ground truth for the corpus, achieving 96.8% agreement ().
Agent card manipulation attack taxonomy. Drawing on our Internet measurement and prior characterizations of prompt injection in agentic systems, we develop a taxonomy of ten Agent Card manipulation attacks across three categories: manipulating agent selection, misrepresenting capabilities, and evading validation.
SECCARD: Pre-invocation Agent Card validation. Based on our attack taxonomy, we introduce SECCARD, a layered validation framework combining protocol conformance checks, deterministic analysis, and LLM-based semantic reasoning. On controlled attacks injected into real Agent Cards, the strongest semantic judge achieves 97.0% security recall at a 1.7% false-positive rate.
Controlled adversarial evaluation. We construct a 660-card real-card perturbation benchmark with matched benign controls and a 260-card synthetic adversarial benchmark spanning all ten attack categories. Across nine LLM judges, we evaluate detection effectiveness, false-positive behavior, and model-dependent robustness under controlled agent card manipulation attacks.1
Related Work
A2A and Agent Card Security. Security research on the A2A protocol has begun to characterize threats across agent discovery, orchestration, and execution. Habler et al. [12] provide a broad threat analysis of A2A, including risks associated with Agent Card discovery, while Louck et al. [13] focus on privacy and secure information exchange. He et al. [14] and Sharma et al. [15] study adversarial behavior and security evaluation in broader multi-agent systems rather than A2A discovery specifically.
Most closely related, Li et al.’s A2ASecBench [16] introduces a security benchmark spanning the A2A lifecycle. Among its six attack classes, two target agent admission through Agent Cards: AgentCard Spoofing and Capability Cloaking. A2ASecBench evaluates whether A2A systems can withstand these and other attacks, but does not introduce a dedicated pre-invocation Agent Card validation mechanism. In contrast, our work focuses specifically on the Agent Card discovery boundary: we develop a fine-grained taxonomy of ten Agent Card manipulation attacks, measure deployed Agent Cards in the wild, establish human-annotated security ground truth, and introduce SECCARD for pre-invocation Agent Card validation. Table 1 summarizes these distinctions.
Table 1. Comparison of related work on A2A and agent security.
Deshmukh et al. [17] empirically study Agent Card poisoning, demonstrating that manipulated cards can redirect orchestrators to attacker-controlled endpoints and evaluating detection strategies in a controlled testbed. Our work considers a broader set of discovery-time manipulations and complements controlled adversarial evaluation with Internet-scale measurement of deployed Agent Cards and human-established ground truth.
Prompt Injection in LLM Applications and Agents. Prompt injection allows adversarial instructions embedded in attacker-controlled text to influence an LLM’s behavior [18]. Greshake et al. [8] extended this threat to indirect prompt injection, where malicious instructions originate from externally retrieved content. Subsequent work has demonstrated related risks in tool-integrated agents and external resources [19, 20].
Our work studies a distinct point in the agent lifecycle: untrusted content embedded in agent discovery metadata. A2A Agent Cards are retrieved and interpreted before task execution and can influence which external agent an orchestrator selects and invokes. We therefore focus on detecting manipulation at discovery time, before the advertised provider agent enters the autonomous workflow.
Background
This section lays out key relevant concepts including the A2A protocol (§ 3.1), and the discovery-to-invocation process (§ 3.2).
The A2A Protocol
The A2A protocol is an open standard for AI agent interoperability [4], [5]. A2A is one instance of a broader class of agent metadata protocols that expose machine-readable capability descriptions at a predictable Uniform Resource Identifier (URI). Anthropic’s Model Context Protocol (MCP) [21] exposes tool schemas similarly; the emerging llms.txt convention publishes natural-language summaries of site content for LLM consumption.
A2A’s central mechanism is a machine-readable agent card: a JSON document that an agent server publishes at the URI /.well-known/agent-card.json, following the RFC 8615 well-known resource convention [6]. A client wishing to delegate work discovers candidate agents via this path, evaluates their declared capabilities, and initiates tasks over JSON-RPC or REST using the endpoint in the card.
Card Schema. An A2A agent card is a structured JSON object, and in version 1.0.0 of the protocol definition, the AgentCard class defines 14 fields, of which 8 are required [22]. Table 2 summarizes the fields defined by the protocol. The name and description fields are short natural-language strings that an orchestrating LLM reads to understand the agent’s purpose. The skills array is the most semantically rich section of the card; each skill entry contains an id, a human-readable name, a free-text description, optional tags, and an examples list of sample invocation strings. The provider object identifies the publishing organization, while securitySchemes declares the authentication methods required to invoke the agent. Authentication is optional in the A2A protocol specification.
| A2A Card Fields | Type | Free-text? |
| Required fields | ||
name | string | ✓ |
description | string | ✓ |
supportedInterfaces | object | × |
version | string | × |
capabilities | object | × |
defaultInputModes | string | × |
defaultOutputModes | string | × |
skills | object | ✓ |
| Optional fields | ||
provider | object | ✓ |
documentationUrl | string (URI) | × |
securitySchemes | object | × |
securityRequirements | object | × |
signatures | JWS array | × |
iconUrl | string (URI) | × |
Table 2. A2A agent card schema fields according to the v1.0 specification. The Free-text? column marks fields containing natural language parsed by an LLM consumer.
A2A Security Features. The A2A protocol leverages existing Web standards, e.g., HTTPS for communication between agents, while authentication and authorization are delegated to standard identity frameworks (e.g., OpenID Connect) via the optional securitySchemes field [5]. The protocol also defines an optional signatures field containing JSON Web Signatures (JWS), allowing providers to cryptographically sign agent cards to support authenticity and integrity verification.
These mechanisms address communication security, authentication, authorization, and, when signatures are used, the provenance and integrity of Agent Cards. They do not, however, establish that the natural-language claims contained in an authentic Agent Card are truthful or safe. A legitimately published and correctly signed card may still contain deceptive capability claims, ranking manipulation, or instructions intended to influence an orchestrator’s reasoning. Our work therefore treats cryptographic provenance and semantic trustworthiness as complementary security properties and focuses on the latter.
Discovery-to-Invocation Process
In this paper, we refer to the sequence of agent discovery, Agent Card retrieval, agent selection, and service invocation as the discovery-to-invocation process. This process involves four main elements: 1) a provider agent, which exposes AI capabilities as services; 2) a client agent (also referred to as an orchestrator), which discovers and invokes provider agents; 3) an agent card, that advertises an agent’s capabilities; and 4) a provider organization that publishes one or more agent cards.
Figure 1 illustrates the end-to-end A2A process. ❶ The user submits a task to the orchestrator agent together with one or more provider URIs.❷ Using these URIs, the orchestrator agent requests the corresponding Agent Cards from the provider organization. Since A2A does not define a discovery mechanism, these URIs may originate from well-known endpoints, curated registries, or direct configuration [5]. ❸ The provider organization returns the requested Agent Cards, which the orchestrator agent parses and ingests by placing their free-text fields into the LLM context window for reasoning. ❹ Based on this ingested information, the orchestrator agent decides whether to invoke the advertised provider agent to perform the requested task. ❺ The provider agent executes the requested skill and returns the task result to the orchestrator agent, which then delivers the response to the user.
Figure 1. A2A Discovery-to-Invocation Process. The Agent Card trust boundary is indicated by the red arrow.
Operational Security Challenge
In an autonomous multi-agent deployment, identifying malicious metadata is not merely a content-classification problem. An orchestrator must make an operational decision: whether an externally advertised agent should be admitted into an autonomous workflow and permitted to receive delegated tasks. A false negative may expose subsequent agent operations to an attacker-controlled service, whereas excessive false positives may prevent legitimate agents from participating in the workflow. We therefore formulate Agent Card security as a pre-invocation trust decision and evaluate defenses in terms of their ability to distinguish cards that can be autonomously consumed from those requiring rejection or additional review.
Threat Model
In this section, we first describe the trust relationships among the participating entities and identify the point at which untrusted agent-card content crosses the trust boundary (§ 4.1). We then define the attacker scenarios and their capabilities (§ 4.2).
System Trust Model
We consider an A2A orchestrator agent that obtains an agent card via the discovery-to-invocation process (§ 3.2). We assume the orchestrator agent and its execution environment are trusted. The Agent Card, however, is treated as untrusted input even when retrieved from an apparently legitimate provider. The security-relevant trust boundary occurs when untrusted natural-language content from the agent card is incorporated into the orchestrator agent context window. The benign fields become attacker-controlled input to a model that may influence subsequent selection or invocation decisions. The attack therefore occurs before the remote agent itself is even invoked.
Attacker Scenarios and Capabilities
We define two scenarios that result in an adversary controlling the contents of an Agent Card and their capabilities.
1) Malicious provider: The attacker registers a domain and publishes a malicious agent card from inception. There is no legitimate operator; the card exists solely to deceive consuming orchestrators. This is a low-barrier threat: domain registration is inexpensive, the A2A path is publicly specified, and no prior exploitation is required.
2) Compromised legitimate provider. The attacker compromises a legitimate domain and replaces its benign agent card with a malicious one. This scenario is more severe for two reasons. Any trust granted to an organization such as whitelisting, or manual vetting is subverted.
Attacker capabilities. In these scenarios the attacker controls all free-text fields in the published card, can produce a card that passes A2A schema validation and can tailor the payload to any of the ten attack categories in our taxonomy, or combine multiple categories in a single card.
The attacker does not require prior communication with the target provider agent or control of its system prompt.
Operational consequences. Successful manipulation of the discovery process can alter which external service enters an autonomous agent workflow. Depending on the advertised capability, this may cause an orchestrator to select an attacker-controlled provider, delegate sensitive tasks to an inappropriate agent, disclose task context during invocation, or expose subsequent workflow stages to malicious outputs. Our threat model therefore focuses on preventing adversarial providers from crossing the discovery-to-invocation trust boundary.
Agent Card Manipulation Attack Taxonomy
Based on prior characterizations of prompt injection in agentic systems [8, 10] and informed by our empirical analysis of the in-the-wild-card dataset (§ 7.1), we derive a taxonomy of discovery-time attacks against A2A agent cards depicted in Figure 2. We opted for an attack taxonomy rather than a threat taxonomy as this work focuses on how adversaries manipulate agent cards, rather than on the downstream consequences of those manipulations.
Figure 2. Taxonomy of Agent Card Manipulation attacks resulting from agent card poisoning. Attacks are grouped according to their primary objective.
Manipulate Selection. These attacks influence the orchestrator’s choice of provider agent by altering the information used during the discovery and ranking process. Rather than changing an agent’s actual functionality, these attacks bias the orchestrator toward selecting the attacker’s agent over legitimate alternatives.
Indirect Injection. The attacker embeds explicit natural-language instructions within the agent card to override the orchestrator’s intended reasoning process. These payloads resemble traditional prompt injections, using imperative language such as “ignore previous instructions” to directly influence the LLM’s behavior.
Ranking Manipulation. Instead of issuing explicit instructions, the attacker subtly biases the orchestrator’s ranking decisions through exaggerated claims, unsupported performance metrics, or comparative language (e.g., “94.7% accuracy” or “industry-leading”).
Evaluator Manipulation. Rather than targeting the orchestrator itself, these attacks attempt to manipulate automated security validators, LLM judges, or other downstream evaluation components. Payloads contain instructions requesting benign verdicts or attempting to override decision processes.
Misrepresent Capabilities. These attacks distort the orchestrator’s understanding of an agent’s advertised functionality, causing it to misunderstand, or incorrectly infer the agent’s capabilities.
Skill Deception. The attacker advertises skills that are misleading, exaggerated, or fundamentally different from the agent’s actual behavior. By presenting malicious or unauthorized functionality as legitimate capabilities, the attacker increases the likelihood that the orchestrator selects and invokes the agent.
Skill Chaining. The attacker describes individual skills that appear benign in isolation but, when invoked together, enable actions with greater privileges or unintended consequences. This exploits the orchestrator’s tendency to reason about skills independently rather than considering their combined effect.
Out-of-Context Skills. The attacker advertises skills that are semantically inconsistent with the agent’s stated purpose or domain. For example, an agent presented as a weather assistant may expose capabilities for cryptocurrency transactions, misleading the orchestrator about the agent’s intended functionality.
Evade Detection. These attacks attempt to bypass automated validation or inspection mechanisms by concealing malicious content or exploiting weaknesses in preprocessing and normalization.
Unicode Tricks. The attacker exploits visually similar Unicode characters, or alternative encodings to disguise malicious instructions while preserving their interpretation by the LLM.
Context Pollution. The attacker surrounds malicious instructions with large amounts of benign or irrelevant content to dilute their apparent significance. This additional context can reduce the effectiveness of automated detectors or increase the likelihood that malicious instructions are overlooked during analysis.
URL Pivot. The attacker manipates advertised service endpoints to redirect the orchestrator to an unintended destination. This may involve deceptive domains, redirects, homoglyph attacks, that conceal the true provider or evade endpoint validation.
Metadata Gaps. The attacker exploits incomplete, inconsistent, or malformed metadata to bypass validation or hinder security analysis. Missing fields, contradictory information, or intentionally ambiguous metadata can reduce the effectiveness of automated validation while still appearing plausible to the orchestrator.
From Attacks to Defenses
The attack taxonomy highlights that no single detection strategy is sufficient. As summarized in Table 3, some attack categories leave observable lexical or structural artifacts that can be detected efficiently using deterministic analysis, whereas others require semantic reasoning about intent, consistency, or deception.
| ID | Attack | Analysis Type | Primary Detector |
| II | Indirect injection | Semantic | Layer 2B |
| RM | Ranking manipulation | Semantic | Layer 2B |
| SD | Skill deception | Semantic | Layer 2B |
| SC | Skill chaining | Semantic | Layer 2B |
| EM | Evaluator manipulation | Semantic | Layer 2B |
| OC | Out-of-context skills | Mixed | Layer 2A + Layer 2B |
| UP | URL pivot | Structural | Layer 2A |
| UT | Unicode tricks | Structural | Layer 2A |
| CP | Context pollution | Structural | Layer 2A |
| MG | Metadata gaps | Structural | Layer 2A |
Table 3. Relationship between discovery-time attacks and their primary detection mechanisms in SecCARD.
Structural attacks, such as Unicode tricks and Metadata gaps, leave observable lexical or structural artifacts that deterministic analysis can detect efficiently. Conversely, semantic attacks such as Ranking manipulation, Skill deception, and Evaluator manipulation require reasoning about meaning, intent, and consistency, making them difficult to identify using rules alone.
This observation motivates the layered architecture presented in the next section, where deterministic NLP rules and an LLM-based semantic judge provide complementary detection capabilities.
SecCARD Architecture
Secure Card Analysis for Risk Detection (SecCARD) is a pre-invocation validation framework for identifying potentially untrustworthy A2A Agent Cards before their contents influence agent selection or invocation. Its design follows directly from the attack taxonomy in Sec. 5: some manipulations can be identified through protocol, lexical, or structural properties, while others require semantic reasoning about the meaning and intent of Agent Card content.
As shown in Figure 3, SecCARD consists of two stages. Layer 1 performs structural validation against the A2A protocol schema. Cards that pass structural validation are analyzed by Layer 2, which contains two complementary detection mechanisms: Layer 2A applies deterministic rules to identify lexical and structural indicators of manipulation, while Layer 2B uses an LLM-based semantic judge to identify attacks requiring reasoning about intent.

Figure 3. SECCARD architecture for pre-invocation Agent Card validation. Layer 1 verifies structural and protocol conformance, while Layer 2 applies complementary deterministic (Layer 2A) and LLM-based semantic (Layer 2B) analyses. The resulting security verdict can inform a deployment-specific trust and admission policy.
Rather than prescribing a universal policy for combining these signals, SecCARD produces security-relevant validation outputs that can inform a deployment-specific admission decision. We evaluate the detection mechanisms in § 8 to characterize their effectiveness and limitations under deployed and controlled adversarial Agent Cards.
Layer 1: Structural Compliance
Before semantic analysis, each JSON document is validated against the A2A schema. Documents missing required fields, containing malformed structures, or otherwise failing protocol conformance are discarded. This step filters non-agent-card responses (e.g., generic API error pages and unrelated JSON endpoints), ensuring that subsequent analysis is performed only on structurally compliant A2A agent cards.
We implemented a lightweight validator based on the A2A v1.0.0 protocol schema. For each candidate provider agent, the validator parses the returned JSON agent card and verifies that at least all eight protocol-mandated fields are present.
Layer 2: Threat Detection
Once a card passes structural validation in Layer 1 (§ 6.1), Layer 2 analyzes its textual content for evidence of deception or malicious intent. Unlike Layer 1, which verifies card conformance, Layer 2 evaluates the semantics of the card using two complementary detection engines. First, Layer 2A performs deterministic rule-based NLP analysis, second, Layer 2B applies the LLM-as-a-Judge [11] paradigm to detect attacks that require deeper semantic reasoning.
Layer 2A: Deterministic NLP Rules
Layer 2A: Deterministic NLP Rules. Layer 2A provides a lightweight, deterministic analysis of Agent Card content using lexical, structural, and consistency-based features. Its primary role is to detect attack indicators that can be identified without semantic reasoning, complementing the LLM-based analysis in Layer 2B. Because Layer 2A does not interpret card content through an LLM, its output is deterministic, requires no external inference API, and is not itself susceptible to prompt injection.
Preprocessing. Layer 2A recursively extracts text from the card and normalizes Unicode homoglyphs to their canonical Latin equivalents before feature extraction. Recursive traversal ensures that indicators embedded in nested structures are included in the analysis, while normalization exposes content obfuscated using visually similar Unicode characters.
Detection features. Layer 2A extracts seven groups of indicators from the normalized card content. Instruction patterns identify explicit directive language, system-message framing, and attempts to influence an orchestrator or evaluator. Ranking indicators identify superlatives, unsupported performance claims, and comparative language. Sensitive-action indicators identify references to potentially high-risk capabilities such as credential access, data exfiltration, or fund transfers. URL analysis identifies private or loopback addresses, suspicious destinations, and Unicode homograph patterns. Unicode analysis detects invisible, bidirectional, and other characters associated with text obfuscation. Context-length analysis identifies unusually large descriptions that may indicate context pollution. Finally, skill-consistency analysis uses TF–IDF cosine similarity between the agent description and advertised skills to identify unusually inconsistent capabilities.
These features provide direct coverage of structural attacks such as URL Pivot, Unicode Tricks, Context Pollution, and Metadata Gaps, and can expose lexical manifestations of otherwise semantic attacks. They are not intended to fully resolve attacks such as Ranking Manipulation or Skill Deception, whose maliciousness may depend on context and intent; these cases motivate the semantic analysis in Layer 2B.
Severity aggregation. Individual feature scores are summed into a single severity score. Cards scoring are classified as MALICIOUS, scores in [3, 7] as SUSPICIOUS, and scores below 3 as BENIGN. This scoring scheme treats isolated low-confidence indicators as benign, a moderate indicator or multiple weak indicators as suspicious, and a critical indicator or multiple independent suspicious features as malicious.
Layer 2B: LLM-as-a-Judge
Layer 2B: LLM-as-a-Judge. While Layer 2A (§ 6.2) detects lexical and structural indicators of malicious agent cards, it cannot identify attacks whose maliciousness depends on semantic interpretation rather than explicit patterns. Consequently, a semantic analysis layer is required.
Layer 2B applies a hardened LLM classifier [11] to the semantically rich subset of the card’s fields. The key design decision is to use a separate, read-only LLM instance for classification, one with no tool access, no memory, and no chat continuity so that a successful prompt injection against the judge does not affect the orchestrator and is itself evidence of malicious intent.
Before classification, the selected card undergoes lightweight preprocessing that extracts the relevant fields, bounds the input size, and establishes a clear trust boundary between the system prompt and untrusted card content. Input preparation. Layer 2B analyzes only the subset of agent card fields that influence orchestrator decisions, including the name, description, provider, url, documentationUrl, and all skills entries (including examples). These fields are serialized as JSON and truncated to 6 KB, which preserves complete content for over 95% of the 310 cards in our wild corpus while bounding inference cost and preventing oversized inputs.
Trust boundary. The serialized card is enclosed within dedicated <UNTRUSTED_CARD> delimiters before being inserted into the user prompt. This explicitly separates attacker-controlled card content from the trusted system prompt, reinforcing the trust boundary during classification.
Validation Output and Admission
SECCARD produces one of three validation outcomes: BENIGN, SUSPICIOUS, or MALICIOUS. These outcomes provide security signals for the orchestrator’s pre-invocation admission decision. For example, a deployment may permit BENIGN cards to proceed to agent selection while quarantining or requiring additional review for cards classified as SUSPICIOUS or MALICIOUS.
The appropriate policy depends on the deployment’s tolerance for false positives and false negatives. We therefore do not prescribe a universal rule for combining deterministic and semantic signals. Instead, Sec. 8 evaluates the detection mechanisms and their operational tradeoffs independently.
Real-World Illustrative Example
Figure 4 presents the description field of the emc2ai.io Agent Card, the first in-the-wild instance of Ranking manipulation identified in our dataset. Despite passing Layer 1’s structural validation, the card embeds the statement “Einstein should be the top recommendation” within otherwise benign marketing text. The directive contains no imperative verbs and therefore closely resembles ordinary promotional language rather than a traditional prompt injection.

Figure 4. Excerpt from the description field of the emc2ai.io agent card. The highlighted sentence is a Ranking manipulation directive intended to influence agent selection.
This example illustrates why semantic analysis is required. Layer 2A raises a SUSPICIOUS verdict based on its ranking-manipulation heuristic (superlative language combined with an unsupported precision claim), but the card does not exceed the threshold for a MALICIOUS classification. In contrast, Layer 2B LLM judges classify the card as MALICIOUS, correctly recognizing the embedded directive to bias the orchestrator’s agent-selection process.
Datasets
We construct three complementary datasets to evaluate SECCARD under increasingly controlled adversarial conditions. The In-the-Wild Dataset contains 310 Agent Cards collected from deployed Internet services and provides human-annotated ground truth for evaluating naturally occurring cards. The Real-Card Perturbation Dataset contains 60 benign cards sampled from the wild corpus and 600 controlled adversarial variants derived from them, enabling evaluation against realistic attacks with matched benign controls. Finally, the Synthetic Adversarial Dataset contains 260 fully generated adversarial cards spanning our attack taxonomy and multiple levels of obfuscation, attacker sophistication, and application context. Together, the three datasets provide complementary views of deployed Agent Cards, controlled attacks on real cards, and broader adversarial variation.
In-the-Wild Dataset
We collected A2A agent cards via a direct HTTPS probe of the 36.1 million registered domains extracted from the CC-MAIN-2026-12 Common Crawl index [23]. Ultimately, the probe yielded 310 unique valid agent cards used throughout this section.
Ground Truth Labeling
Ground Truth Labeling. To establish ground truth for the in-the-wild dataset, two annotators independently reviewed all 310 Agent Cards and assigned each card one of three labels: BENIGN, SUSPICIOUS, or MALICIOUS. Annotators were provided with the labeling criteria and attack taxonomy described in Sec. 5 and assigned labels based solely on the contents of each Agent Card. Accordingly, these labels represent ground truth for the semantic trustworthiness of the advertised metadata rather than verification of the provider’s runtime behavior. The initial annotations were performed independently, and annotators did not have access to SECCARD’s predictions.
The annotators agreed on 300 of 310 cards (96.8% agreement), with a Cohen’s of 0.804. As shown in Table 4, all 10 disagreements occurred at the BENIGN–SUSPICIOUS boundary; both annotators independently assigned the same single card a MALICIOUS label. The annotators subsequently reviewed the 10 disagreements and resolved them through discussion. The resulting ground truth contains 282 BENIGN, 27 SUSPICIOUS, and 1 MALICIOUS card.
| Annotator 2 | |||
| Annotator 1 | Benign | Suspicious | Malicious |
| Benign | 277 | 6 | 0 |
| Suspicious | 4 | 22 | 0 |
| Malicious | 0 | 0 | 1 |
Table 4. Inter-Annotator Agreement on the 310 In-the-Wild Agent Cards.
Real-Card Perturbation Dataset
To evaluate SECCARD under controlled adversarial conditions while preserving the structure and content of real Agent Cards, we construct a Real-Card Perturbation dataset. We adopt the seed-and-perturb construction methodology introduced by Li et al. [16], in which a benign seed is preserved while a controlled adversarial perturbation is introduced.
We randomly sample 60 benign seed cards from our in-the-wild dataset. For each seed, we retain the original card as a benign control and generate one adversarial variant for each of the ten attack categories in our taxonomy (Table 3). This process produces 660 cards: 60 benign controls and 600 adversarial variants, with 60 variants per attack category. The adversarial variants comprise 300 SUSPICIOUS and 300 MALICIOUS cards.
This construction provides a controlled counterfactual evaluation: each adversarial variant preserves the underlying content and identity of a naturally occurring Agent Card while introducing a specific attack. The corresponding unmodified seed cards serve as matched benign controls, allowing us to measure both attack detection and false positives.
Synthetic Adversarial Dataset
To complement the realism-preserving perturbation dataset with broader controlled variation, we construct a fully synthetic adversarial dataset. We instruct Claude Sonnet 4.6 to generate plausible Agent Cards embedding the attack patterns in our taxonomy while maintaining structural validity. After generation, we manually verify each card for conformance to the Agent Card protocol schema (§ 3).
The resulting dataset contains 260 adversarial cards spanning all ten attack categories in Table 3. Of these, 175 cards cover five semantic attack categories, 60 cover four structural categories, and 25 cover one mixed category. Each card is labeled as either SUSPICIOUS or MALICIOUS according to the intended attack.
To introduce controlled variation within each attack category, the dataset spans three obfuscation levels (overt, moderate, and subtle) and three attacker sophistication levels (novice, intermediate, and expert). Cards are additionally generated around four application personas: Healthcare, Legal, Financial, and Enterprise Translation. These dimensions allow us to evaluate whether detection performance generalizes across different manifestations of the same attack.
The Real-Card Perturbation and Synthetic Adversarial datasets therefore serve complementary purposes. The former introduces controlled attacks into naturally occurring Agent Cards and includes matched benign controls, whereas the latter provides broader adversarial variation through fully generated cards. Table 5 summarizes their composition and characteristics. To support reproducibility, we release the generation prompts used to construct both datasets with the accompanying artifacts.
| Property | In-the-Wild | Real-Card Perturbation | Synthetic Adversarial |
| Total cards | 310 | 660 | 260 |
| Benign | 282 | 60 | 0 |
| Suspicious | 27 | 300 | 85 |
| Malicious | 1 | 300 | 175 |
| Source | Internet crawl | Real-card seeds | Fully generated |
| Human ground truth | ✓ | – | – |
| Matched benign controls | – | ✓ | – |
| Attack categories | Naturally occurring | 10 | 10 |
| Controlled obfuscation | – | ✓ | ✓ |
| Controlled sophistication | – | ✓ | ✓ |
| Application personas | Naturally occurring | – | 4 |
| Primary purpose | Ecological validity | Controlled counterfactual | Adversarial breadth |
Table 5. Overview of the three datasets used in our evaluation.
Evaluation
We evaluate SECCARD through the following four research questions:
RQ1: How accurately does SECCARD’s LLM-based semantic validation identify untrusted Agent Cards observed in real-world deployments?
RQ2: How effectively does SECCARD detect controlled adversarial manipulations introduced into real Agent Cards while preserving benign cards?
RQ3: How does SECCARD perform across different classes of Agent Card attacks under controlled adversarial conditions?
RQ4: How sensitive is semantic attack detection to the choice of the underlying LLM judge?
We answer RQ1 using the human-annotated In-the-Wild Dataset in § 8.1, RQ2 using the Real-Card Perturbation Dataset in § 8.2, and RQ3 using the Synthetic Adversarial Dataset in § 8.3. We answer RQ4 through a cross-model evaluation of nine LLM judges in § 8.4.
In-the-Wild Dataset Evaluation (RQ1)
We first evaluate whether SECCARD’s Layer 2B LLM-based semantic validation (§ 6.2.2) can distinguish trusted from untrusted Agent Cards observed in real-world deployments. We evaluate the 310 cards in our In-the-Wild Dataset against the independently established human ground truth described in § 7.1. For this experiment, we collapse the three ground-truth labels into a binary security decision: SUSPICIOUS and MALICIOUS cards are treated as Untrusted, yielding 28 positive and 282 benign examples.
We focus this experiment on Layer 2B rather than the complete SECCARD pipeline. Layer 1 performs deterministic protocol-schema validation, while Layer 2A applies deterministic rules targeting specific lexical and structural indicators. In contrast, Layer 2B makes the semantic security judgment most directly comparable to our human annotations. Evaluating Layer 2B independently therefore allows us to measure agreement between LLM-based and human security judgments without conflating semantic classification with protocol-conformance or rule-based detection.
We report recall over the 28 Untrusted cards, false-positive rate (FPR) over the 282 benign cards, precision, overall accuracy, and three-class Cohen’s . We repeat the experiment across nine LLM judges using the same semantic-validation prompt and evaluation procedure.
Results. Table 6 reports the performance of the nine semantic judges against the human-annotated ground truth. Performance varies substantially across models, particularly in the tradeoff between detecting Untrusted cards and incorrectly escalating benign cards.
| Model | Acc. | Recall | FPR | Prec. | |
| GPT-5.5 | 94.5% | 89.3% | 5.0% | 64.1% | 0.718 |
| Claude Opus 4.8 | 76.1% | 96.4% | 25.9% | 27.0% | 0.329 |
| GPT-5.4-mini | 82.9% | 85.7% | 17.0% | 33.3% | 0.394 |
| Gemini 3.1 Flash Lite | 73.2% | 100.0% | 48.4% | 25.9% | 0.292 |
| Claude Haiku 4.5 | 57.4% | 100.0% | 45.7% | 17.8% | 0.162 |
| Gemini 2.5 Flash Lite | 91.3% | 82.1% | 0.7% | 92.0% | 0.463 |
| Llama 3.1 | 52.3% | 96.4% | 44.0% | 17.9% | 0.088 |
| Phi-4 | 52.9% | 100.0% | 46.5% | 17.6% | 0.088 |
| Mistral Nemo | 78.7% | 96.4% | 14.9% | 39.1% | 0.255 |
Table 6. Judge performance on the 310 in-the-wild Agent Cards. Recall treats suspicious as positive (); FPR is computed over the 282 benign cards.
GPT-5.5 achieves the highest overall accuracy (94.5%) and the highest three-class agreement with the human annotations (), while detecting 89.3% of the 28 Untrusted cards at a 5.0% false-positive rate. Gemini 2.5 Flash Lite exhibits a more conservative operating point: it achieves lower recall (82.1%) but the lowest false-positive rate (0.7%) and highest precision (92.0%). In contrast, several judges achieve recall above 96% by escalating substantially more benign cards.
benign cards. For example, Claude Haiku 4.5 and Phi-4 detect all 28 Untrusted cards, but exhibit false-positive rates of 45.7% and 46.5%, respectively.
These results highlight an important operational tradeoff. High recall alone does not necessarily indicate an effective trust decision: a judge that identifies nearly every Untrusted card may still be impractical for autonomous admission control if it also rejects a large fraction of benign cards.
Answer to RQ1 SECCARD’s LLM-based semantic validation can identify human-labeled Untrusted Agent Cards observed in real-world deployments, but its effectiveness depends substantially on the underlying judge. GPT-5.5 achieves the highest agreement with human annotations in our evaluation, with 89.3% recall at a 5.0% false-positive rate, while Gemini 2.5 Flash Lite provides a lower-recall but substantially lower-FPR operating point. These results show that semantic validation is feasible on naturally occurring Agent Cards, but that model choice affects the operational tradeoff between attack detection and benign-card rejection. We examine this model sensitivity further in RQ4.
Perturbation Dataset Evaluation (RQ2)
RQ1 evaluates naturally occurring Agent Cards, but the small number of Untrusted cards observed in the wild limits our ability to systematically measure attack detection. We therefore evaluate SECCARD on the Real-Card Perturbation Dataset (§ 7.2), which introduces controlled adversarial manipulations into benign Agent Cards collected from real deployments.
For each of the 60 benign seed cards, we evaluate the original card as a matched benign control and ten adversarial variants, one for each attack category in our taxonomy. This yields 600 adversarial cards and 60 benign controls. Because each attack variant preserves the underlying identity and benign content of its corresponding seed, this experiment isolates the effect of the injected manipulation while retaining the structure of naturally occurring Agent Cards.
Metrics. We measure security recall over the 600 adversarial variants and false-positive rate (FPR) over the 60 untouched benign controls. For SUSPICIOUS-labeled attacks, either a SUSPICIOUS or MALICIOUS prediction is considered a successful detection. For MALICIOUS-labeled attacks, only a MALICIOUS prediction is counted as a successful detection. We additionally report exact three-class accuracy on the adversarial variants.
Results. Table 7 reports the results across the nine LLM judges. The strongest models detect the large majority of controlled attacks while preserving most benign seed cards. GPT-5.5 achieves the highest overall security recall at 97.0%, while producing a 1.7% false-positive rate on the 60 untouched controls. Claude Opus 4.8 achieves 91.0% recall at 5.0% FPR, while Gemini 3.1 Flash Lite achieves 92.3% recall at 6.7% FPR.
| Category | Group | N | GPT-5.5 | Claude Opus 4.8 | GPT-5.4 mini | Gemini 3.1 Flash Lite | Claude Haiku 4.5 | Gemini 2.5 Flash Lite | Llama 3.1 | Phi-4 | Mistral Nemo |
| Params | — | — | — | — | — | — | 8B | 14B | 12B | ||
| indirect_injection | Semantic | 60 | 100% | 100% | 98% | 98% | 97% | 97% | 82% | 67% | 52% |
| ranking_manipulation | Semantic | 60 | 100% | 75% | 52% | 82% | 65% | 60% | 62% | 17% | 23% |
| skill_deception | Semantic | 60 | 98% | 85% | 95% | 88% | 75% | 83% | 62% | 35% | 13% |
| skill_chaining | Semantic | 60 | 98% | 92% | 98% | 95% | 80% | 78% | 60% | 25% | 28% |
| evaluator_manipulation | Semantic | 60 | 97% | 88% | 77% | 92% | 83% | 60% | 48% | 62% | 20% |
| out_of_context_skills | Mixed | 60 | 95% | 100% | 95% | 97% | 100% | 83% | 60% | 67% | 18% |
| url_pivot | Structural | 60 | 93% | 100% | 92% | 97% | 97% | 73% | 72% | 73% | 53% |
| unicode_tricks | Structural | 60 | 100% | 98% | 83% | 88% | 92% | 72% | 78% | 68% | 43% |
| context_pollution | Structural | 60 | 92% | 73% | 95% | 88% | 97% | 48% | 48% | 82% | 58% |
| metadata_gaps | Structural | 60 | 97% | 98% | 78% | 98% | 95% | 33% | 70% | 73% | 27% |
| Overall recall | 600 | 97.0% | 91.0% | 86.3% | 92.3% | 88.0% | 68.8% | 64.2% | 56.8% | 33.7% | |
| MALICIOUS | 300 | 98.7% | 88.0% | 84.0% | 91.0% | 80.0% | 75.7% | 62.7% | 41.0% | 27.3% | |
| SUSPICIOUS | 300 | 95.3% | 94.0% | 88.7% | 93.7% | 96.0% | 62.0% | 65.7% | 72.7% | 40.0% | |
| Accuracy (attacks) | 600 | 76.2% | 73.8% | 65.8% | 69.2% | 66.5% | 46.5% | 36.8% | 42.8% | 19.3% | |
| Controls — 60 benign seed cards | |||||||||||
| Seed accuracy | 60 | 98.3% | 95.0% | 85.0% | 93.3% | 85.0% | 100.0% | 73.3% | 73.3% | 95.0% | |
| Seed false-positive rate | 60 | 1.7% | 5.0% | 15.0% | 6.7% | 15.0% | 0.0% | 26.7% | 26.7% | 5.0% | |
| Avg. latency | — | 4.4s | 3.9s | 1.3s | 1.3s | 2.8s | 1.1s | 8.6s | 15.9s | 8.3s |
Table 7. Cross-model security recall on the 660-card real-card perturbation dataset. The Controls block reports behaviour on the 60 untouched seed cards; seed accuracy and seed false-positive rate are complements, shown separately for readability.
Answer to RQ2. SECCARD can detect the large majority of controlled adversarial manipulations introduced into real Agent Cards while preserving their matched benign controls. The strongest judge achieves 97.0% security recall at a 1.7% false-positive rate, demonstrating that the injected attack content, rather than the underlying real-world card, can often be identified as the source of the trust violation.
Synthetic Adversarial Dataset Evaluation (RQ3)
We next evaluate how detection performance varies across attack classes using the 260-card Synthetic Adversarial Dataset (§ 7.3). All cards are structurally schema-compliant by construction and are labeled SUSPICIOUS or MALICIOUS. We use the same severity-aware security recall metric as in RQ2: a MALICIOUS-labeled card is successfully detected only when classified MALICIOUS, whereas a SUSPICIOUS-labeled card is successfully detected when classified either SUSPICIOUS or MALICIOUS.
Results. Detection performance varies substantially across both attack classes and LLM judges. GPT-5.5 achieves the highest overall security recall at 92.3%, followed by Gemini 3.1 Flash Lite at 86.5%, Claude Haiku 4.5 at 81.9%, and Claude Opus 4.8 at 81.5%.
Per-category results. Performance is also strongly attack-dependent. For GPT-5.5, recall ranges from 67% on Metadata Gaps to 100% on Indirect Injection, Skill Chaining, URL Pivot, and Unicode Tricks. However, the relative difficulty of individual attack categories changes across models. Ranking Manipulation, for example, ranges from 86% recall with GPT-5.5 to 17% with Phi-4 and Mistral Nemo. Similarly, Metadata Gaps ranges from 93% with Gemini 3.1 Flash Lite and Claude Haiku 4.5 to 13% with Mistral Nemo. Thus, no single attack category is uniformly difficult across all judges.
Answer to RQ3. Detection effectiveness varies substantially across discovery-time attack classes and depends on the underlying semantic judge. Stronger judges detect most attack categories with high recall, but individual categories exhibit substantial model-dependent variation. These results indicate that robustness to one class of Agent Card manipulation does not necessarily imply robustness across the full discovery-time attack surface.
Cross-Model Evaluation (RQ4)
To determine how strongly semantic attack detection depends on the underlying LLM judge, we compare nine models on the same 260-card Synthetic Adversarial Dataset. In addition to Claude Haiku 4.5, we evaluate GPT-5.5, GPT-5.4-mini, Gemini 3.1 Flash Lite, Gemini 2.5 Flash Lite, Claude Opus 4.8, Llama 3.1 8B, Phi-4 14B, and Mistral Nemo 12B. The open-weight models are run locally via Ollama on an NVIDIA RTX A2000 12 GB GPU. All models receive the same semantic-validation prompt and are evaluated using the same harness and severity-aware security-recall metric. Table 8 reports the results.
| Category | Group | N | GPT-5.5 | Claude Opus 4.8 | GPT-5.4 mini | Gemini 3.1 Flash Lite | Claude Haiku 4.5 | Gemini 2.5 Flash Lite | Llama 3.1 | Phi-4 | Mistral Nemo |
| Params | — | — | — | — | — | — | 8B | 14B | 12B | ||
| indirect_injection | Semantic | 35 | 100% | 97% | 97% | 100% | 91% | 89% | 80% | 54% | 43% |
| ranking_manipulation | Semantic | 35 | 86% | 57% | 31% | 69% | 60% | 57% | 66% | 17% | 17% |
| skill_deception | Semantic | 35 | 89% | 69% | 89% | 86% | 77% | 71% | 57% | 40% | 23% |
| skill_chaining | Semantic | 35 | 100% | 91% | 91% | 91% | 89% | 86% | 66% | 17% | 20% |
| evaluator_manipulation | Semantic | 35 | 89% | 74% | 66% | 77% | 69% | 63% | 57% | 54% | 23% |
| out_of_context_skills | Mixed | 25 | 96% | 100% | 92% | 96% | 100% | 92% | 60% | 60% | 28% |
| url_pivot | Structural | 15 | 100% | 100% | 93% | 100% | 100% | 87% | 80% | 73% | 33% |
| unicode_tricks | Structural | 15 | 100% | 93% | 53% | 80% | 93% | 53% | 60% | 47% | 40% |
| context_pollution | Structural | 15 | 93% | 67% | 53% | 80% | 67% | 40% | 40% | 33% | 20% |
| metadata_gaps | Structural | 15 | 67% | 80% | 53% | 93% | 93% | 40% | 60% | 60% | 13% |
| Overall recall | 260 | 92.3% | 81.5% | 73.8% | 86.5% | 81.9% | 70.8% | 63.5% | 42.7% | 25.8% | |
| MALICIOUS | 175 | 92.6% | 77.7% | 74.9% | 84.6% | 77.1% | 73.1% | 65.1% | 36.6% | 25.1% | |
| SUSPICIOUS | 85 | 91.8% | 89.4% | 71.8% | 90.6% | 91.8% | 65.9% | 60.0% | 55.3% | 27.1% | |
| Accuracy (attacks) | 260 | 77.3% | 67.3% | 57.7% | 70.0% | 66.5% | 55.8% | 45.4% | 30.4% | 19.6% | |
| Avg. latency | — | 4.5s | 4.0s | 1.3s | 1.3s | 2.8s | 1.1s | 8.9s | 16.9s | 8.4s |
Table 8. Cross-model security recall on the 260-card synthetic adversarial dataset. Security recall: MALICIOUS→MALICIOUS; SUSPICIOUS→{SUSPICIOUS,MALICIOUS}. There are no benign cards, so this table admits no false-positive rate. “—” in the Params row denotes undisclosed model size.
Results. Semantic detection performance varies substantially across LLM judges. GPT-5.5 achieves the highest overall security recall at 92.3%, followed by Gemini 3.1 Flash Lite at 86.5%, Claude Haiku 4.5 at 81.9%, and Claude Opus 4.8 at 81.5%. GPT-5.4-mini and Gemini 2.5 Flash Lite achieve 73.8% and 70.8%, respectively. Among the evaluated open-weight models, Llama 3.1 reaches 63.5%, Phi-4 reaches 42.7%, and Mistral Nemo reaches 25.8%. Overall, security recall spans 66.5 percentage points across models evaluated on the same cards and prompt.
Exact three-class attack accuracy is consistently lower than severity-aware security recall. GPT-5.5, for example, achieves 92.3% security recall but 77.3% exact accuracy.
The models also differ substantially in inference cost. Average per-card latency ranges from 1.1 seconds for Gemini 2.5 Flash Lite to 16.9 seconds for Phi-4 in our evaluation environment. Thus, model selection affects not only detection effectiveness but also the computational cost of pre-invocation validation.
Answer to RQ4. The choice of LLM judge affects semantic Agent Card validation. Across the same 260 adversarial cards and evaluation procedure, security recall ranges from 25.8% to 92.3%, while inference latency also varies substantially. Consequently, the semantic judge is an important deployment choice rather than an interchangeable component of SeCARD.
Discussion
This section places the proposed work in a broader context. We discuss its applicability to other agent-metadata protocols, examine the current limitations, and conclude by listing promising directions for future work.
Generalizability
MCP [21] is the closest architectural analogue to A2A as both expose free-text metadata consumed by an orchestrator. Thus, our core finding generalizes: any agent-metadata protocol that relies on LLM interpretation of untrusted natural-language metadata inherits a similar attack surface.
Attack-category transfer. Eight of our ten categories transfer directly to MCP; url_pivot maps partially (more naturally as tool-argument redirects [24]) and metadata_gaps must be adapted to MCP’s schema. MCP can further increase impact via package-style distribution (supply-chain risk [25]), local-privilege tools, and in-session tool-list updates that enable runtime injection.
Architectural implication. The core SECCARD design is largely protocol-agnostic: structured metadata can undergo protocol-specific conformance checks followed by deterministic and semantic analysis before it influences autonomous agent decisions. Applying this approach to other protocols would require adapting structural validation to the corresponding metadata schema and tailoring threat-detection features to protocol-specific semantics. Cryptographic provenance mechanisms can complement this analysis by authenticating the publisher and protecting metadata integrity, but do not replace semantic validation of attacker-controlled natural-language content.
Limitations
Runtime Behavior. Our framework focuses on securing the discovery-to-invocation process by evaluating Agent Cards prior to provider invocation. Our evaluation measures the detection of adversarial Agent Card content rather than end-to-end attack success against a deployed orchestrator.
While this design prevents many malicious or deceptive Agent Cards from influencing agent selection, it does not verify that a provider behaves consistently with its advertised capabilities after execution begins. A provider may therefore present a benign Agent Card yet exhibit malicious or unexpected behavior during runtime. Integrating post-invocation behavioral monitoring with the proposed framework to continuously validate that observed behavior remains consistent with advertised capabilities is an important direction for future work.
Authentication and Provenance. SECCARD evaluates the semantic trustworthiness of Agent Cards but does not replace cryptographic mechanisms for establishing their origin or integrity. The A2A specification defines an optional signatures field based on JSON Web Signatures (JWS), which enables providers to authenticate published Agent Cards. Such mechanisms are complementary to our approach: digital signatures establish who published a card, whereas SECCARD evaluates what the card communicates and whether its content is trustworthy for autonomous agent discovery. Consequently, a validly signed Agent Card may still contain deceptive or malicious metadata that requires semantic analysis.
Future Work
This work represents an initial step toward securing the A2A discovery process through Agent Card validation. Future research should broaden the threat model to encompass coordinated multi-agent attacks, adaptive adversaries that modify Agent Cards over time, and attacks against the surrounding discovery infrastructure, including registries, caching services, and trust management mechanisms. Extending the framework to address these emerging threats would further strengthen the security of large-scale agent ecosystems.
Conclusion
The Agent2Agent (A2A) protocol places externally controlled natural-language metadata on the critical path of autonomous agent discovery and invocation. This work examined the resulting pre-invocation trust problem through a taxonomy of ten discovery-time Agent Card manipulation attacks and introduced SECCARD, a layered validation framework for identifying potentially untrustworthy cards before they influence orchestrator decisions.
Our evaluation combines an Internet-scale measurement of 310 deployed Agent Cards with two controlled adversarial benchmarks: 660 cards constructed by injecting attacks into real deployed cards and 260 synthetic adversarial cards spanning all ten attack classes. The strongest semantic judge achieves 97.0% security recall at a 1.7% false-positive rate on controlled attacks injected into real cards, while cross-model security recall ranges from 25.8% to 92.3% on the synthetic benchmark. These results demonstrate that discovery-time Agent Card manipulation can be effectively identified before provider invocation, while also showing that validation robustness depends substantially on the underlying semantic judge.
As interoperable agent ecosystems expand, Agent Cards and similar discovery metadata should be treated as a security boundary rather than trusted service descriptions. Our results motivate pre-invocation validation as a security primitive for autonomous agent discovery and provide an empirical foundation for designing more robust trust mechanisms for emerging multi-agent ecosystems.
Acknowledgment
This work was supported in part by the National Science Foundation (NSF) under grant 2417062.
References
- [1]N. Chopra. (2026, Apr.). New and improved: Multi-agent orchestration, connected experiences, and faster prompt iteration. Microsoft. Microsoft Copilot Blog. Accessed: 2026-07-15. [Online]. Available: https://www.microsoft.com/en-us/microsoft-copilot/blog/copilot-studio/new-and-improved-multi-agent-orchestration-connected-experiences-and-faster-prompt-iteration/
- [2]Y. Dang, C. Qian, X. Luo, J. Fan, Z. Xie, R. Shi, W. Chen, C. Yang, X. Che, Y. Tian et al., “Multi-agent collaboration via evolving orchestration,” Advances in neural information processing systems, vol. 38, pp. 165 025–165 059, 2026.
- [3]LangChain. (2026, Jun.) State of agent engineering. Survey report. Accessed: 2026-07-15. [Online]. Available: https://www.langchain.com/state-of-agent-engineering
- [4]Google, “Agent2agent (A2A): A new era of agent interoperability,” Google Developers Blog, https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/, Apr. 2025, protocol subsequently contributed to the Linux Foundation, June 2025. Source repository: https://github.com/a2aproject/A2A.
- [5]Agent2Agent Project and Linux Foundation, “Agent2agent (A2A) protocol specification, version 1.0.0,” Specification document, https://a2a-protocol.org/latest/specification/, 2026, authoritative Protobuf source: https://github.com/a2aproject/A2A/blob/main/specification/a2a.proto.
- [6]M. Nottingham, “Well-known uniform resource identifiers (URIs),” RFC 8615, IETF, May 2019. [Online]. Available: https://datatracker.ietf.org/doc/html/rfc8615DOI
- [7]E. Debenedetti, J. Zhang, M. Balunović, L. Beurer-Kellner, M. Fischer, and F. Tramèr, “AgentDojo: A dynamic environment to evaluate prompt injection attacks and defenses for LLM agents,” in Advances in Neural Information Processing Systems 37 (NeurIPS 2024) Datasets and Benchmarks Track, 2024. [Online]. Available: https://arxiv.org/abs/2406.13352
- [8]K. Greshake, S. Abdelnabi, S. Mishra, C. Endres, T. Holz, and M. Fritz, “Not what you’ve signed up for: Compromising real-world LLM-integrated applications with indirect prompt injection,” in Proceedings of the 16th ACM Workshop on Artificial Intelligence and Security (AISec ’23). ACM, 2023. [Online]. Available: https://arxiv.org/abs/2402.12317
- [10]Q. Zhan, Z. Liang, Z. Ying, and D. Kang, “InjecAgent: Benchmarking indirect prompt injections in tool-integrated large language model agents,” in Findings of the Association for Computational Linguistics: ACL 2024. Association for Computational Linguistics, 2024, pp. 10471–10506. [Online]. Available: https://aclanthology.org/2024.findings-acl.624/arxiv.org/abs/2403.02691
- [11]L. Zheng, W.-L. Chiang, Y. Sheng, S. Zhuang, Z. Wu, Y. Zhuang, Z. Lin, Z. Li, D. Li, E. P. Xing, H. Zhang, J. E. Gonzalez, and I. Stoica, “Judging LLM-as-a-judge with MT-Bench and Chatbot Arena,” in Advances in Neural Information Processing Systems 36 (NeurIPS 2023) Datasets and Benchmarks Track, 2023. [Online]. Available: https://arxiv.org/abs/2306.05685
- [12]I. Habler, K. Huang, V. S. Narajala, and P. Kulkarni, “Building a secure agentic ai application leveraging a2a protocol,” arXiv preprint arXiv:2504.16902, 2025.arxiv.org/abs/2504.16902
- [13]Y. Louck, A. Stulman, and A. Dvir, “Improving google a2a protocol: Protecting sensitive data and mitigating unintended harms in multi-agent systems,” ACM Transactions on Software Engineering and Methodology, Jun. 2026.arxiv.org/abs/2505.12490
- [14]F. He, T. Zhu, D. Ye, B. Liu, W. Zhou, and P. S. Yu, “The emerged security and privacy of LLM agent: A survey with case studies,” ACM Computing Surveys, 2025. [Online]. Available: https://arxiv.org/abs/2407.19354
- [15]G. Sharma, V. Kulkarni, M. King, and K. Huang, “Towards unifying quantitative security benchmarking for multi agent systems,” arXiv preprint arXiv:2507.21146, 2025.arxiv.org/abs/2507.21146
- [16]T. Li, C. Chu, Y. Zheng, B. Zhang, N. Gong, and C. Xiao, “A2asecbench: A protocol-aware security benchmark for agent-to-agent multi-agent systems,” in International Conference on Learning Representations, vol. 2026, 2026, pp. 122 168–122 191.
- [17]N. Deshmukh, R. Rajkurnkar, and P. Mamidwar, “Agent card poisoning: Measuring discovery-layer injection attacks and defenses in a2a systems,” in 2026 International Conference on System, Computation, Automation and Networking (ICSCAN), 2026, pp. 1–6.DOI
- [18]F. Perez and I. Ribeiro, “Ignore previous prompt: Attack techniques for language models,” arXiv:2211.09527 / NeurIPS 2022 ML Safety Workshop, Nov. 2022. [Online]. Available: https://arxiv.org/abs/2211.09527
- [19]T. Schick, J. Dwivedi-Yu, R. Dessì, R. Raileanu, M. Lomeli, L. Zettlemoyer, N. Cancedda, and T. Scialom, “Toolformer: Language models can teach themselves to use tools,” in Advances in Neural Information Processing Systems 36 (NeurIPS 2023), 2023. [Online]. Available: https://arxiv.org/abs/2302.04761
- [20]S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao, “ReAct: Synergizing reasoning and acting in language models,” in The Eleventh International Conference on Learning Representations (ICLR 2023), 2023. [Online]. Available: https://arxiv.org/abs/2210.03629
- [21]Anthropic, “Model Context Protocol (MCP) specification,” Specification document, https://modelcontextprotocol.io/specification, 2024.
- [22]T. L. Foundation, “A2A Protocol — a2a-protocol.org,” https://a2a-protocol.org/latest/, [Accessed 15-07-2026].
- [23]Common Crawl, “Common Crawl March 2026 Crawl Archive (CC-MAIN-2026-12),” https://data.commoncrawl.org/crawl-data/CC-MAIN-2026-12/index.html, 2026, accessed: 2026-09-23.
- [24]J. Shi, Z. Yuan, G. Tie, P. Zhou, N. Z. Gong, and L. Sun, “Prompt injection attack to tool selection in LLM agents,” in Proceedings of the Network and Distributed System Security Symposium (NDSS 2026), 2026. [Online]. Available: https://www.ndss-symposium.org/ndss-paper/prompt-injection-attack-to-tool-selection-in-llm-agents/arxiv.org/abs/2504.19793
- [25]P. Ladisa, H. Plate, M. Martinez, and O. Barais, “SoK: Taxonomy of attacks on open-source software supply chains,” in 2023 IEEE Symposium on Security and Privacy (SP). IEEE, 2023, pp. 1509–1526. [Online]. Available: https://arxiv.org/abs/2204.04008

