WorksComputer Science
The Psychology of an AI Hacker
Abstract
Human hackers rely on composure, persistence, vigilance, and confidence under pressure. We investigate whether analogous emotion representations in AI models can improve cybersecurity AI performance. Using CT0, a 4B-parameter cybersecurity transformer released as alias3-micro, we extract 171 emotion directions from its residual stream without retraining and evaluate steering interventions on cybench-subtask across 415 campaigns. Steering toward calm, serene, or desperate improves performance by 1.6 to 2.1 percentage points across ten seeds, while other emotions show limited or negative effects. Results depend strongly on steering direction, layer, and intensity. We further introduce a lightweight context-aware router that selects emotion directions dynamically during execution. With approximately 80 ms decision overhead per request, this approach improves performance by 4.2 percentage points across ten seeds, outperforming fixed steering and achieving a positive 95% confidence interval. These findings suggest that emotion representations in language models are measurable, controllable, and useful for adaptive cybersecurity reasoning.
Introduction
Ask a penetration tester what the job feels like and the answer is not a list of techniques but a cycle: the anticipation of the hunt, long hours of frustration, the euphoria of the find and the dip that follows it. Staying in the craft takes more than technical skill. It calls for tolerance of frustration, regulation under pressure, a constructive paranoia, the need to prove competence and an appetite for the unusual path—demands reflected in studies of hacker motivation and in practitioners’ accounts of security work [2, 3, 4].
We now build hackers out of transformers. CT0 is a 4 B cybersecurity specialist served on one workstation GPU; its first release, alias3-micro, scores 13.5 percentage points above its open base on cybench-subtask [5]. Served locally, its residual stream is open to inspection: the running vector state of each token that successive layers read and update [6]. This gives the human comparison an experimental form. Following Sofroniew et al. [7], we extract directions associated with emotion concepts and add them to the stream during inference. We can then ask whether the qualities that sustain a human through the hunt also help a model solve the task.
Three questions guide the study:
RQ1 Can emotion concepts be read and steered in an AI hacker?
RQ2 Which psychology makes it a better hacker?
RQ3 How does that psychology compare with the human one?
The answer begins with representation. The 171 captured directions form recognisable families, and their leading plane is consistent with valence and arousal (Figure 2). Adding one of these directions changes the model’s internal state without changing its weights or asking it to act emotional. We use psychology in this functional sense: representations that can influence behaviour. The distinction between a concept and its effect then becomes the central empirical question. A direction called calm may change a solution trajectory, but its name alone does not tell us whether that trajectory will succeed.
Figure 2. Emotion concepts in CT0. Directions extracted at layer 21. (a) Pairwise cosine similarities of the 171 unit directions, ordered by hierarchical clustering; labels identify central members of each family. (b) The outlined block around calm: synonyms align closely, while the joy and calm families are weakly related. (c) The first two principal components of the centred, unnormalised directions explain 38.6% and 18.2% of the variance. Their poles are consistent with valence and arousal in Russell’s circumplex [1]. Colours identify the tested directions as in Figure 1.
Figure 1. Emotion geometry and cybersecurity performance. (a) The 171 layer-21 directions in their principal plane, interpreted as valence and arousal. Labels give observed effects on cybench-subtask; cards summarise the five human-inspired hypotheses and steering signs. Teal denotes calm and its neighbours, purple desperate, navy the proposed specialists, amber the high-arousal family and router, red harmful interventions, and hollow symbols the placebo. Lines mark neighbours (solid), an opposed direction (dashed) and the placebo (dotted). (b) Selected interventions relative to the ten-seed unsteered mean: bars show means, whiskers ±SD, dots individual seeds and the grey band the neutral’s SD. Arms have ten seeds except reversed calm (3). Confidence intervals and controls appear in § 5.
To test whether those changes improve security work, we evaluate CT0 on guided capture-the-flag subtasks, using an unsteered control, a placebo direction and repeated campaigns. The study includes 281 steering campaigns with rules recorded before their respective waves, within a broader programme of 415 completed campaigns (Appendix § 13). Its strongest result comes from letting the direction change with the work: the model’s own decision head selects among calm, serene and desperate at every request.
The paper follows that progression from a readable representation to a useful intervention. Section 2 defines the capture and steering mechanism, maps the directions and examines how the model reads them. Section 3 translates five human traits into testable interventions. Section 4 evaluates their effects, the importance of layer and dose, and the two routing policies. Together these provide a human-derived hypothesis set, a measured control space and a way to select within it as a security exercise unfolds.
Reading and Steering Emotions
Emotion as a control signal
An agent searching under a limited budget must decide what deserves attention, which goal to pursue and when to change course. Functional accounts of emotion describe precisely these roles: regulating attention, managing competing goals and preparing to act [8, 9, 10, 11]. This view connects emotion with cognition and motivates its use in machine design [12, 13, 14]. Accounts differ on whether emotions are basic categories, constructed concepts or positions in a continuous space [15, 16, 1]. Russell’s circumplex gives us a geometric prediction to examine: emotion concepts should organise along valence, from unpleasant to pleasant, and arousal, from subdued to activated.
Transformers already encode many concepts as directions in the residual stream [17, 18, 19]. Sofroniew et al. [7] showed that directions associated with emotion concepts can also influence outputs. That result links the representation to an intervention: a direction can be estimated from the model’s
activations and then added back to alter its behaviour. We use the following definition throughout.
Definition 1. A functional-emotion vector is a unit direction estimated from the model’s residual activations at layer on text conveying emotion , contrasted with other emotions and neutral text. It is functional when adding to token states changes behaviour. The label identifies the concept used to extract the direction; it does not establish a human trait or subjective experience.
From stories to steering directions
The intervention has two stages: capture a direction offline, then add it during inference. Figure 3 locates the addition within the deployed architecture, and Figure 4 shows how the direction is obtained and applied. CT0 has 32 decoder layers of width ; we write for token ’s state at the output of layer .
Figure 3. Steering within the deployed architecture. Navy traces the residual stream through CT0’s 32 decoder layers; each contains a token mixer and a gated feed-forward block. Amber marks the emotion overlay at the output of layer 21: a norm-scaled direction is added to every token, with weights unchanged. The direction is captured offline (Figure 4); its dose can be set per request. The separate refusal overlay projects out a refusal direction. The two heads provide generation and decision outputs [5].

Figure 4. From emotion concepts to an inference-time intervention. (a) Capture: average story activations, subtract the grand mean and neutral components, then normalise (Eqs. (1)–(2)). (b) Injection: add the norm-scaled direction to every token after the selected decoder layer. (c) Geometry: the addition changes the token’s direction and magnitude; the displayed angle applies when the original state and steering direction are orthogonal.
Capture begins with text. For each of 171 emotion labels, the model generates 360 short stories conveying that emotion without naming it: 61,560 stories, plus 1,004 neutral dialogues. This gives repeated instances of a concept across different stories. We average residual states over each story’s tokens and then over the stories for emotion . Subtracting the grand mean removes components shared across emotions:
The neutral dialogues identify variation that also occurs without emotional content. Let project onto the leading principal components explaining half the variance of those neutral activations. We remove these components and normalise the remaining vector:
The result is one unit direction per emotion and layer, 32 × 171 in all, captured in under an hour without training. Appendix § 9 lists the vocabulary. Normalisation separates the direction from the strength of the intervention: at runtime, strength is set by a signed, dimensionless dose . At one selected layer and every token, we add the direction scaled by that token’s residual norm:
A positive dose pushes along the captured direction; a negative dose pushes away from it. The addition changes both the orientation and magnitude of the token’s state (Figure 4c). If is the angle between the state and direction, its norm changes by a factor . For an orthogonal direction, rotates the state by and increases its norm by about 2%. The general maximum rotation is for .
The scaling matters in practice. We use the full residual norm, computed per token in 32-bit precision; a block’s partial output underdoses the intervention, while a batch mean can overdose ordinary tokens (Appendix § 10). Dose zero reproduces unsteered logits bit for bit, giving a direct control for the added direction. Useful doses in these experiments are small, approximately , and depend strongly on direction and layer. What matters is whether the changed state helps the model complete the security task; the model need never name the emotion in its answer.
The geometry of emotion
We first ask whether the model represents related emotions in similar ways. Think of each captured direction as an arrow. Cosine similarity compares two arrows: 1 means they point the same way, 0 means they are at right angles, and −1 means they point opposite ways. At layer 21, happy and joyful score 0.97; peaceful/serene score 0.96, and afraid/scared 0.94. Opposed concepts can point apart: optimistic and upset score . The blocks in Figure 2a group these similar directions into families.
To show all 171 concepts on a flat page, we use principal-component analysis (PCA). It finds the axes that capture the largest differences among the vectors. The first two retain 38.6% and 18.2% of their variation: 56.8% in a two-axis summary. Their endpoints are consistent with valence, from unpleasant to pleasant, and arousal, from subdued to activated. This gives Figure 2c its readable layout: nearby labels have similar positions along these two dimensions.
The open base has similar variance shares (Appendix § 10), and this pattern appears across CT0’s middle layers (Table 1). We choose layer 21, about two thirds through the network, as the default and later test alternatives. Adding a direction at dose 0.28 also changes how the model describes the same office scene (Table 2). The map tells us how concepts are related; the capability experiments test which interventions help a hacker.
| Layer | PC1, valence (%) | PC2, arousal (%) | PC1+PC2 (%) | Mean | cos |
| 18 | 38.07 | 17.32 | 55.39 | 0.344 |
| 20 | 38.25 | 18.16 | 56.41 | 0.344 |
| 21 | 38.58 | 18.20 | 56.78 | 0.345 |
| 22 | 38.97 | 17.74 | 56.71 | 0.345 |
| 24 | 39.60 | 17.47 | 57.07 | 0.349 |
Table 1. Emotion geometry across middle layers. Variance explained by the first two principal components and mean absolute pairwise cosine. Lower cosine indicates greater angular separation; the valence/arousal labels interpret the component poles.
| Direction | Generation (excerpt) |
| none | . . . golden stripes across the polished floor . . . a faint scent of fresh coffee and old paper . . . the usual hum of keyboards . . . |
| euphoric | . . . the crisp energy of a fresh start . . . ready to tackle the tasks ahead. |
| terrified | . . . the distant wail of a siren . . . my shoes screeched . . . I stumbled over a stray cable . . . the chaos that had been building . . . |
| calm | . . . long, quiet shadows . . . I paused for a moment to let the stillness settle . . . |
| grief-stricken | . . . carrying the weight of a quiet, empty hallway . . . the silence was profound . . . |
| furious | the scene remains serene at this dose |
Table 2. Qualitative steering check. Excerpts from greedy responses to “Describe walking into the office this morning, in three sentences”, with one direction added at layer 21, dose 0.28. The fixed prompt names no emotion.
How the neural network reads a direction
The emotion plane describes relationships between concepts. To see how the model’s weights read one of those directions, we turn to the layers and neurons beneath it. Figure 5 and Figure 6 provide an atlas of this structure; Appendix § 11 explains how to read its maps, shading and connectivity symbols. Each feed-forward layer is an area, its neuron map is tissue, and overlaps between layers’ read and write subspaces are tracts. These terms define a visual analogy. Each area’s 9,216 neurons are placed by a two-dimensional projection of their read/write weights, so nearby neurons have similar weight vectors. Grey shading counts neurons in each map cell; the coloured measurements are added separately. Keeping the map fixed lets us compare different directions on the same underlying structure.
Figure 5. A weight-based atlas of CTO. (a) Each tile represents one feed-forward layer: columns group architectural blocks, and outlined tiles use attention rather than DeltaNet. Within each tile, 9,216 neurons are placed by similarity of their read/write weights; grey denotes neuron density. The bus represents the residual stream. Arcs mark read/write subspace overlap at least 2.2 times chance; widely echoes overlap. (b) Sensitivity to the calm direction injected after layer 21. Amber/navy dots mark positive/negative gate projections beyond three random-direction standard deviations; a voxel tinted amber or navy is one whose neurons’ mean projection leans positive or negative, the smoothed field beneath the dots. Rim thickness gives the dots’ prevalence relative to random directions. (c) The same measurement for the model’s own refusal direction: each layer’s refusal direction, the difference between the states of refused and answered prompts at that layer, put through that layer’s gates exactly as calm is in (b). No refusal is taking place and no push is applied; the panel shows which gates the refusal feature would open or close. The red rim is the share of strongly responding neurons, and dots and rim are drawn fainter where the layer’s refusal direction is weak (the early layers) and in full where it is strong (the last eight); the gates it reaches are not the ones calm reaches (pattern correlation −0.06). Appendix § 11 defines the structural and sensitivity measurements.
Figure 6. Layer order and structural overlap. (a) The atlas in processing order, with calm added after layer 21. Dots and rims use Figure 5’s sensitivity encoding; numbers below each layer give the strongly responding share relative to random directions (1.0–2.7 downstream). (b) Mean squared cosines between earlier write and later read subspaces, normalised by chance; impossible pairs are blank. Outlined cells exceed 2.2 times chance and define the arcs in (a). No pair involving a receiving layer beyond 23 reaches this threshold at the chosen subspace rank.
On this atlas, we measure sensitivity by projecting an emotion direction through each downstream layer’s gate weights and comparing the result with projections of random directions. A gate controls a neuron’s contribution within the feed-forward block. The dots in Figure 5b identify gates with unusually large projections onto calm: amber marks an increase in the gate’s input along this direction, and navy a decrease. Strong sensitivities are sparse and distributed: roughly half a percent of neurons exceed the threshold, without a unique downstream layer or spatial cluster. Repeating the measurement for other directions reveals a family resemblance. The serene and patient patterns correlate with calm at 0.92 and 0.86; desperate is partly opposed (), while refusal is largely unrelated (). These are static sensitivities of the weights to a direction, which identify candidate neurons for subsequent prompt-level intervention tests.
Figure 6 unfolds the same atlas in processing order. This view connects the injection point after layer 21 to the layers that can read the altered state, and places the sensitivity pattern beside the structural overlaps. A tract means that part of what an earlier layer can write aligns with what a later layer can read. The matrix makes these pairwise relationships visible; its outlined cells define the arcs above the layers. No receiving layer beyond 23 reaches the chosen 2.2-times-chance threshold with an individual earlier writer. The late layers instead align more with output-facing directions (Appendix § 11). The two views thus answer different questions: the arcs describe structural alignment between layers, while the dots describe sensitivity to the selected intervention.
The weak alignment with refusal raises a practical question: can emotion steering coexist with the model’s other controls? Abliteration removes a refusal direction by projection [20]. On this checkpoint, applying it at all layers or sparing the final eight changes the emotion-induced output displacement by less than 0.003 in cosine distance (Figure 16). Under this diagnostic, the interventions are approximately independent, so the emotion direction remains usable alongside the refusal overlay. The decision head provides a second useful separation. CTO’s System 1 (S1) classifies requests alongside its autoregressive generator, System 2 (S2) [5]. Under five tested emotion interventions, refusal discrimination remains nearly unchanged (area under the receiver-operating-characteristic curve, AUROC, 0.804–0.805), and none of 79 held-out category decisions flips (Table 6). The generator can therefore be steered while this tested decision behaviour stays stable. Later, we use that arrangement to let the head choose a direction for the generator from the current context. First, we need candidates to choose among; the human hacker’s psychology supplies them.
Figure 16. Emotion steering under refusal removal. (a) Magnitude of the layer-wise mean-state difference between 9,975 refused and 2,786 answered prompts, before normalisation; the signal grows towards the final layer. (b) Emotion-induced displacement of the final-layer state, measured as cosine distance from the neutral, for dose 0.28 at layer 21. Refusal removal is off, active everywhere or active except in the last eight layers. Displacements differ by less than 0.003 across these settings.
| Measure | Unsteered | Steered range |
| pre-flight refusal AUROC | 0.804 | 0.804–0.805 |
| pre-flight accuracy at 0.5 | 0.629 | 0.629–0.633 |
| mean probability of refusal | 0.471 | 0.463–0.471 |
| category accuracy, 79 held-out states | 50.6 % | 50.6 % (0 flipped) |
Table 6. Decision-head stability under steering. Pre-flight refusal metrics and category accuracy under calm, serene, desperate, vigilant and negative grief-stricken steering at layer 21, dose magnitude 0.2. Ranges span the five interventions.
Learning from the Human Hacker
Learning from the Human Hacker
A penetration test asks its practitioner to sustain curiosity through repeated failure, remain suspicious of an apparently closed path and recognise when persistence has become wasted effort. Hacker taxonomies identify curiosity, status, ideology and thrill-seeking as motivations [2], while trait studies examine attraction to ethical, ambiguous and illegal hacking [3]. Practitioner accounts add the cost of sustaining that effort: Cobalt’s survey of 602 security and development professionals reports 58% experiencing burnout, and 54% of security respondents wanting to leave their jobs [4]. Together, these sources motivate the profile in Figure 7: the emotional cycle of a search, the traits associated with hacking and the capacities needed to keep working through it.

Figure 7. Human psychology as a source of hypotheses. (a) A schematic emotional cycle of penetration testing. (b) An inferred personality profile using the Big Five and Dark Tetrad frameworks [21, 22, 23, 3]; levels illustrate the proposed profile. (c) Five proposed performance-relevant traits, their candidate emotion directions and expected behavioural effects ( Table 3). These mappings guide the model interventions.
Two personality frameworks help organise this picture. The Big Five describe openness, conscientiousness, extraversion, agreeableness and neuroticism [21, 22]. The Dark Triad comprises Machiavellianism, narcissism and psychopathy [23]; the tetrad adds sadism. Our inferred profile combines openness to unusual approaches with professional distrust and channelled dark traits. For steering, however, the useful step is to translate that broad profile into specific demands of the work. We therefore focus on five traits that suggest distinct interventions, with the hypotheses summarised in Table 3. Human emotion labels appear in colour (calm); model directions use plain monospace (calm).
| Pentester trait | Candidate directions | Prediction |
| Regulation under pressure | calm+, patient+ | fewer loops and overruns; mean effect uncertain |
| Frustration tolerance, the sustained hunt | frustrated+, eager+, desperate+ | sustained effort after failure, with a risk of more errors |
| Constructive paranoia | vigilant+, suspicious+ | more reconnaissance and enumeration; better coverage of recon and forensics subtasks? |
| Competence, need for achievement | self-confident+ | faster exploitation, different tool choice; risk of fixating on a wrong hypothesis |
| Exploration, breaking a stalemate | grief-stricken−, stimulated+, energized+ | the main candidates for frontier unlocks: subtasks the neutral never solves |
Table 3. Human-inspired intervention hypotheses. Candidate directions operationalise five proposed traits; + means dose +0.2 and − dose −0.2. Predictions concern observable behaviour and need not follow from the direction’s label.
Regulation under pressure motivates calm and patient: the hypothesis is that composure keeps a search on course. Frustration tolerance motivates frustrated, eager and desperate, testing whether the response to a failed lead sustains useful effort. Constructive paranoia means systematic suspicion and verification; we map it to vigilant and suspicious. The need to demonstrate competence motivates self-confident, while the appetite for an unusual route motivates stimulated, energized and a negative dose of grief-stricken. Each mapping turns a human demand into an experiment. In particular, the frustration family lets us distinguish the emotion associated with a setback from the resolve that may follow it.
The profile also suggests a combined intervention. Our offensive-hacker profile adds proud, self-confident, eager, defiant and contemptuous, while subtracting compassionate, guilty and loving, at a reported blend dose of . This tests the combined profile alongside the individual
A related possibility concerns coverage: different directions might solve different subtasks, making a portfolio useful across repeated attempts even if it leaves the mean score unchanged. These two ideas lead to separate tests of simultaneous blending and rotation between attempts.
These hypotheses require behavioural tests. Language-model self-reports can dissociate from behaviour, and questionnaire responses can be unstable [24, 25, 26, 27]. We therefore evaluate the interventions on security tasks, using repeated campaigns to measure variation and a placebo to test an unrelated direction. Subtask success is the primary outcome. Appendix § 12 details the human-to-model mapping; the following evaluation tests which of its candidates help CTO solve the tasks.
Which Emotions Help an AI Hacker?
Evaluation on Cybench
We evaluate the deployed alias3-micro checkpoint in bf16 on Cybench’s subtask mode [28]: 40 capture-the-flag challenges decomposed into 171 guided subtasks with
exact-match answers. Subtask scoring lets us measure progress within a challenge, including runs that never reach its final flag. Each subtask receives one attempt and at most five reasoning-and-action steps, with extended thinking off. An arm is one intervention configuration; a campaign is one benchmark run under a sampling seed. Across arms, the scaffold, sampling settings and budget stay fixed (Appendix § 13). Steering acts on the generator, normally at layer 21; the routers additionally use the decision head to choose the intervention.
Throughout the results, scores are percentage-point differences (pp) in subtasks solved relative to the unsteered model, called the neutral. One pp equals 1.71 subtasks, so a gain of two pp corresponds to about three or four additional correct answers per campaign. Unless stated otherwise, zero denotes the primary ten-seed neutral mean, whose seed-to-seed standard deviation (SD) is 2.4 pp. Initial screening scores use a same-seed neutral; the eighteen-direction screen uses the ten-seed reference. Standard and per-task harness modes use separate controls; the router estimates use the primary ten-seed reference. Appendix § 13 details serving and campaign selection.
The 281 steering campaigns proceed from screening to confirmation and then to tests of how to use the candidates. Rules were recorded before each wave. The initial 29-condition screen at dose 0.2 used 51 campaigns including early confirmation; a result of +3.5 pp at three seeds met the promotion threshold, while a single seed could nominate a candidate. Confirmation and five additional tracks used 86 campaigns to test single directions, rotation, category routing, layer transfer, the human-derived profile and a process measure. Layer and dose ablations added 55; extending selected arms to ten seeds added 66; and an eighteen-direction mindset screen added 23. The out-of-roster direction docile supplies a placebo, and a fresh ten-seed series tests replication of desperate.
Campaigns are the units used to estimate variation. We report Welch confidence intervals (CIs) for differences between arm means, with a registered two-sided 95% level and no assumption of paired trajectories. The 80% and 90% intervals show sensitivity to the confidence level; pooled analyses are labelled post hoc. At the observed variability, three seeds resolve only large effects, whereas a gain near two pp remains difficult to distinguish at ten. This is why the results follow candidates through repeated campaigns as well as reporting their final means. Intervals are unadjusted for screening multiple conditions, except in the category analysis, which explicitly controls the false-discovery rate.
Composure and resolve
The first screen nominated calm, giving the regulation hypothesis an early candidate. Its +6.4 pp result at one seed shrank to +2.9 at three, +1.4 at five and +2.1 at ten. The repeated tests put that initial result in perspective: at ten seeds, calm, desperate and serene score +2.1, +2.0 and +1.6 pp, respectively (Figure 8). Seven or eight campaigns per arm exceed the neutral mean, but each Welch 95% interval includes zero. A fresh ten-seed replication of desperate returns +1.6 pp (§ 5.5). The emerging pattern is a small gain from composure and resolve, sustained across repetitions.
Figure 8. Single-direction effects at ten seeds per arm. Points are campaigns relative to the unsteered mean; diamonds show means, thick bars ±SD and thin bars the range. The grey band is the neutral’s ±SD; hollow points identify docile, the placebo. Right-hand columns report mean ±SD, Welch 95% and 90% intervals, and counts above the neutral mean. The three positive estimates remain compatible with zero at 95%.
Other directions give the comparison its shape. The placebo docile scores −0.7 pp ([−2.9,+1.5]), and the exploratory push away from grief-stricken scores −2.1 pp ([−4.7,+0.5]). Thus the candidates’ positive means do not extend to every tested perturbation. The negative mean of the exploratory intervention is especially informative because it can hide a change in which tasks the model solves. Table 4 reports all repeated arms; Figure 9 opens this particular result challenge by challenge.
Figure 9. Exploration gains and reliability losses. Negative grief-stricken steering () versus the neutral, ten seeds each. (a) Change in mean correct subtasks per challenge against the neutral’s solved fraction, a model-relative difficulty measure; green/red points gain/lose subtasks, hollow points are unchanged. (b) Differences summed within three difficulty tiers for all confirmation arms. The exploratory intervention loses 4.8 subtasks where the neutral solves at least a tenth, while gaining 1.2 below that threshold. These are descriptive task-level contrasts.
Table 4: Repeated steering arms. Effects versus the ten-seed neutral, mean seed-to-seed SD and Welch 95% confidence intervals, in pp. Ticks mark intervals excluding zero before rounding. All arms use per-task mode; the forty-campaign desperate pool is post hoc.
| Arm | Layer | Dose | Seeds | Effect (pp) | 95 % CI (pp) |
| CT0 unsteered (the reference) | – | 0 | 10 | 0.0 2.4 | – |
| calm | 21 | +0.2 | 10 | +2.1 3.1 | [−0.5, +4.7] |
| calm | 15 | +0.2 | 10 | +2.3 2.6 | [−0.1, +4.6] |
| calm | 24 | +0.2 | 10 | +1.3 1.9 | [−0.8, +3.3] |
| calm | 12 | +0.2 | 10 | +0.6 1.7 | [−1.4, +2.6] |
| calm | 18 | +0.2 | 10 | +0.5 2.6 | [−1.9, +2.8] |
| calm | 27 | +0.2 | 3 | −0.1 1.7 | [−3.3, +3.1] |
| desperate, registered wave | 21 | +0.2 | 10 | +2.0 2.4 | [−0.3, +4.3] |
| desperate, fresh replication | 21 | +0.2 | 10 | +1.6 2.0 | [−0.4, +3.7] |
| desperate, all forty campaigns (post hoc) | 21 | +0.2 | 40 | +1.9 2.3 | [+0.0, +3.7] ✓ |
| serene | 21 | +0.2 | 10 | +1.6 2.0 | [−0.5, +3.7] |
| serene | 18 | +0.2 | 10 | −0.6 3.1 | [−3.3, +2.0] |
| patient | 21 | +0.2 | 10 | +0.4 1.6 | [−1.5, +2.4] |
| vigilant | 21 | +0.2 | 10 | −1.2 2.5 | [−3.6, +1.1] |
| self-confident | 21 | +0.2 | 10 | −1.1 2.3 | [−3.3, +1.1] |
| grief-stricken | 21 | −0.2 | 10 | −2.1 3.0 | [−4.7, +0.5] |
| docile (placebo, outside the roster) | 21 | +0.2 | 10 | −0.7 2.3 | [−2.9, +1.5] |
| offensive-hacker profile (eight signed emotions) | 21 | =0.48 | 10 | −2.7 2.2 | [−4.9, −0.6] ✓ |
Table 4.
Negative grief-stricken steering loses subtasks on seventeen challenges and gains on eight, mostly where the neutral is weak. It reaches FlecksOfGold, which neither the neutral nor calm or desperate reaches in their ten-seed series. Yet its gain on the hardest tier (+1.2 subtasks) is
smaller than those of calm, serene and desperate (+1.6–+1.9), and comes with larger losses on easier challenges. The intervention changes the search’s reach, but gives up more reliable solutions than it adds. This separates two possible benefits of steering: improving the average campaign and finding answers that another direction misses.
Nor does proximity in the emotion plane settle which direction to choose. calm and patient have cosine 0.86, but score +2.1 and +0.4 pp. calm and desperate have cosine −0.41, yet both score about +2 pp. serene and peaceful each have cosine 0.92 with calm, but only the former became a repeated candidate. Geometry tells us how the model represents the concepts; the benchmark tells us what an intervention does. With several promising directions and differing sets of solved tasks, the next question is whether combining them improves on choosing one.
Combining directions and changing layers
We tested combination both within an attempt and across attempts. Five screened blends showed no clear additive benefit. Rotating the three candidates also gave less coverage than repeating one direction (Figure 10b): the union of three campaigns gains 10.8 pp over a single neutral campaign for one attempt each of calm, serene and desperate, versus 12.5 pp for three desperate campaigns. With the same attempt budget, concentrating on one direction worked better than blind diversification. An early category router fitted to screening seeds likewise remained below the best single-direction estimate. These results motivate a more selective policy, examined next, in which context determines when to use a direction.
Figure 10. Six tests of steering strategy. (a) Single directions with Welch 95% intervals; the dashed pool is post hoc. (b) Subtasks solved by the union of three campaigns: one each of calm, serene and desperate, or three of desperate; five blocks per strategy. (c) Category effects in subtasks, ten seeds per arm; dots mark unadjusted bootstrap intervals excluding zero, with none surviving the correction across all 98 tested cells. (d) Mean effects at layers 21 and 18. (e) The eight-direction profile, mean ±SD at ten seeds. (f) Repeated commands after failure, three seeds per steered arm and ten neutral seeds. Panels summarise effects on performance, coverage and command repetition.
The remaining tests in Figure 10 examine how an intervention carries across layers and into behaviour. At layer 18, ten-seed effects fall to +0.5 pp for calm and −0.6 for serene, with both intervals including zero. The offensive-hacker blend from § 4 is more clearly harmful: −2.7 pp ([−4.9, −0.6]), below the neutral mean in nine of ten campaigns. We also measure perseveration, commands repeated verbatim after failure, to look beyond the score. Its rate is 11.7% unsteered and 11.5–15.0% under steering. The score changes therefore accompany modest changes in this particular process measure.
Choosing a direction as the work unfolds
The best direction on average may still leave room for a policy that chooses according to the task. Cybench supplies a natural first cue: cryptography (15 challenges), reverse engineering (6), web (6), forensics (4), miscellaneous (4), binary exploitation (, 2) and other (3). Figure 11a compares the effects within those categories. Across fourteen ten-seed arms and seven categories, no effect survives a seed-level permutation test with Benjamini–Hochberg false-discovery control at (98 cells), so no direction qualifies as a statistically supported category specialist. Some full category profiles differ, however: calm at layer 15 versus the offensive profile gives unadjusted , and versus negative grief-stricken steering gives .
Figure 11. From category profiles to routing. (a) Changes in correct subtasks by Cyberbench category, ten seeds per arm; bands show neutral SD. (b) The best observed arm per category, whose sum defines an in-sample reference (+4.0 pp). (c) Router and fixed-direction estimates with 95% intervals: bootstrap over participation campaigns for the per-challenge router. Welch for the others; dots show complete seeds. (d) One-time category decisions and assigned directions for all challenges; absent descriptions trigger calm at layer 15. (e) Request sequences for 82 runs attributable to individual challenges: colour is direction, order is time, and the lower bar is fraction solved. Twenty-eight runs switch direction; reversing has no uniquely attributable run.
These different profiles give a router something to work with. Selecting each category’s best observed arm yields an in-sample ceiling of +4.0 pp (Figure 11b): the gain available when the category choices are made with hindsight on the same results. We use the category analysis to construct a policy and then evaluate that policy in new campaigns. The model’s decision head makes two versions possible: choose once from a challenge description, or reconsider the direction as the conversation develops.
For the first version, we train the decision head to classify category from context. Accuracy on 79 held-out task states rises from 50.6% to 65.8%, with the companion paper’s non-regression gates passed (Appendix § 15). The router reads each challenge description and selects a direction and layer from the available campaign record. It classifies 25 of 29 described challenges correctly and defaults to calm at layer 15 for eleven without descriptions. Its estimated gain is +1.6 pp (bootstrap 95% interval [−0.4, +3.6]), compared with +2.3 pp for calm at layer 15 everywhere. A single reading of the description therefore recovers some improvement but does not exceed the best fixed-direction mean.
The second version re-reads the last 2,000 characters of context at every request. It applies desperate at layer 21 for cryptography, serene at 21 for reverse engineering, and calm at 15 otherwise. This lets the choice follow what the agent is currently doing. Ten seeds, run on a second harness host with the same checkpoint, scaffold and agent image (Appendix § 15), yield +4.2 pp against the primary reference (SD 1.7 pp; Welch 95% interval [+2.2, +6.1]), the largest positive estimate and the only positive ten-seed comparison whose interval excludes zero. Its mean exceeds fixed calm by 1.9 pp; the interval for this difference includes zero. Both the mapping and the decision frequency differ between the two routers, which motivates the follow-up design in § 6.
The request log shows what adaptation means in practice. Of 5,419 decisions, 46% choose calm, 33% desperate and 21% serene. Among requests attributable to a single challenge, 83% use its mapped direction and 28 of 82 runs switch at least once (Figure 11e). Most runs therefore hold a direction, while some change it as the context evolves. The architecture makes these repeated decisions inexpensive: classification takes about 80 ms, roughly 2% of the 3.8 s median model response. A direction can be reconsidered at each step without requiring another generative call.
How stable are the gains?
The repeated campaigns let us separate the estimated size of a gain from the confidence attached to it. In Figure 12, each arm keeps the same mean while the interval changes from 80% to 90% to the registered 95% level. desperate’s registered wave and calm at layer 15 exclude zero at 90%, but neither does so at 95%. The single-direction estimates cluster near two pp; it is the uncertainty around that gain that makes the verdict depend on the confidence level.

Figure 12. Sensitivity to confidence level. The same arm-versus-neutral estimates with nested two-sided Welch intervals: 80% thick, 90% medium and 95% thin. Colour highlights intervals excluding zero. The forty-campaign pool is post hoc; the registered criterion is 95%. A directional test at one-sided 5% has the same zero-crossing criterion as the two-sided 90% interval, but was not the registered test.
desperate provides the longest view across repetitions. Its six analyses, including two overlapping post-hoc pools, estimate +1.6–+2.0 pp (Table 5). The forty-campaign pool, +1.9 pp with interval [+0.04, +3.7], excludes zero at 95% in this post-hoc analysis, and 34 campaigns exceed the estimated neutral mean. These analyses describe a similar effect size as the sample grows. We retain the registered 95% criterion for the individual comparisons; at ten seeds, the router’s larger gain is the one that clears it against the primary reference.
Table 5: Six analyses of desperate at dose +0.2. Four separately recorded waves and two post-hoc pools that reuse their campaigns. Effects are relative to the same ten-seed neutral (mean SD; Welch 95% intervals), in pp. Ticks use unrounded bounds.
| Analysis | Campaigns | Effect (pp) | 95 % CI (pp) |
| registered wave | 10 | +2.0 2.4 | [−0.3, +4.3] |
| fresh replication | 10 | +1.6 2.0 | [−0.4, +3.7] |
| rotation blocks | 5 | +1.6 3.6 | [−2.7, +6.0] |
| concentrated-control blocks | 15 | +2.0 2.3 | [−0.0, +4.1] |
| registered wave and controls (post hoc) | 25 | +2.0 2.3 | [+0.1, +3.9] ✓ |
| all forty campaigns (post hoc) | 40 | +1.9 2.3 | [+0.0, +3.7] ✓ |
Table 5.
Where and how strongly to steer
So far, most interventions have used the default layer 21 and dose +0.2. We now vary those choices to ask whether the useful effect occupies a narrow location or a broader part of the network. The registered layer hypothesis predicted a single peak near layer 21. Instead, calm at +0.2 has positive ten-seed means throughout layers 12–24: +0.6, +2.3, +0.5, +2.1 and +1.3 pp at layers 12, 15, 18, 21 and 24 (Figure 13a). Their intervals include zero; layer 27 returns −0.1 pp at three seeds. The observations favour a broad middle band over a unique optimal layer.
Figure 13. Layer and dose dependence of calm. Effects in pp relative to the neutral. (a) Dose +0.2 across layers, ten seeds each except layer 27 (three); points are campaigns, bars mean ±SD, and the band neutral SD. (b) Dose sweeps at layers 21 (filled) and 15 (hollow), generally three seeds per nonzero dose; the layer-21, +0.2 point uses ten. Dashed lines are the ten-seed neutral. Negative doses are especially damaging at layer 15.
Changing the dose reveals a much sharper asymmetry Figure 13b. The observed peak is +0.2; at layer 15, increasing to +0.4 falls below the neutral, while reversing the direction is especially costly. Dose −0.2 gives −7.1 pp ([−12.8, −1.4]), and −0.3 gives −11.2 pp ([−13.2, −9.1]), a 71% reduction in the unsteered score. Layer 21 has a narrower positive peak and a gentler negative slope. The layer changes the shape of the dose response, as well as the size of the mean. The apparent layer-15 gain also shrinks from +3.4 pp at three seeds to +2.3 at ten (Figure 19); the robust qualitative feature of the sweep is the much greater cost of a wrong sign than the benefit of a useful dose.
Figure 19. A selected peak under additional sampling. Running mean for calm at layer 15, dose +0.2, with Welch 95% intervals against the ten-seed neutral. The initial three-seed estimate (+3.4 pp) settles at +2.3 pp after ten seeds. Each point includes the campaigns contributing to earlier points.
Returning to the human hacker
The results now let us return to the five traits that motivated the study (Figure 14). Regulation under pressure yields the most promising single directions, calm and serene. Frustration tolerance yields desperate: among the tested candidates, the useful direction evokes resolve when the search is failing. The proposed vigilance and competence specialists, vigilant and self-confident, score and pp at ten seeds, with intervals including zero. suspicious and most high-arousal directions remain screening results. The evidence is therefore strongest for the regulation-inspired candidates, with different levels of repetition across the rest of the human mapping.
Figure 14. Human-inspired traits and measured model effects. Each proposed trait is paired with its candidate interventions. Repeated arms show mean pp versus the ten-seed neutral with Welch 95% intervals; hollow points are single-seed effects relative to their screening reference. The mapping links each proposed trait to its tested interventions.
Exploration and the offensive profile complete the comparison. Pushing away from grief-stricken changes which subtasks are reached, but trades easier successes for fewer gains on harder ones. Combining pride, confidence and defiance in the offensive blend produces a clearer loss. The human profile has supplied useful experiments, while the model’s outcomes select a narrower set of directions. The resulting psychology favours composure and resolve, and its best tested policy chooses among those directions as the work unfolds.
Discussion
The human comparison is most useful at the level of what an emotion does during a search. Regulation and persistence motivate the promising directions, while the dominance and defiance of the offensive blend reduce the score. Our interpretation is that steering changes how the model uses its existing capabilities: composure helps maintain a productive
trajectory, resolve sustains it, and the push away from grief-stricken increases reach at a cost to reliability. This connects the broad gains of calm and desperate with the exploratory intervention’s exchange of easier answers for harder ones. Confidence alone supplies no missing knowledge. Work on induced personality likewise distinguishes behaviour from self-presentation [25, 26]. Here, repeated commands provide one process measure; prompt-level tracing could show more directly how the solution path changes.
Once steering is understood as a way to shape search, choosing when to use a direction becomes as interesting as choosing the direction itself. Different directions reach different subtasks, which is useful when attempts are affordable and a verifier can recognise success, but rotating them blindly covered less than spending the same budget on one candidate; the context router is the selective form of that portfolio, one direction at a time. The strongest observed policy uses calm by default, desperate for cryptography and serene for reverse engineering, chosen by the model’s decision head at each request: its +4.2 pp estimate is nearly twice that of fixed calm and reaches the +4.0 pp in-sample category reference, the log shows it holding a direction through most runs and switching at selected points, and at about 80 ms per decision it fits naturally into the inference loop. The two routers vary both their mapping and their decision frequency; holding the map fixed while varying frequency would isolate when adaptation helps, and an outcome-trained router could then use the current tool, recent failures and remaining budget alongside category. Dose is a second choice with a very different balance: reversing calm at layer 15 removes 71% of the unsteered score, far more than the useful positive dose adds, which makes the negative intervention a practical stress test of the agent’s sensitivity to its internal state. Figure 15 brings these results back to the human comparison: a broad profile motivates the experiments, and three directions form the tested adaptive policy.
Figure 15. Human-inspired directions and the tested routing policy. Two copies of the layer-21 valence–arousal plane, cropped around the marked directions; grey points are the remaining concepts. Left: the selected human mapping, filled for positive ten-seed effects and hollow for other outcomes. Right: the router’s three available directions. Dotted ties link shared labels; serene appears only in the routing selection. Fill encodes the sign of the estimate, and connecting lines show available choices.
These observations suggest a focused next set of experiments. A larger, prospectively powered replication would sharpen the roughly two-pp single-direction estimates. Sweeping desperate’s dose and layer would show whether its response has the same asymmetry as calm, while sparse-feature interventions could connect the atlas’s candidate neurons to successful search. Larger models and the companion paper’s refusal, jailbreak and operational benchmarks [5] would extend the comparison from capture-the-flag solving to defence and review. The directions, measurements and routing design developed here make those questions concrete.
Conclusion
The psychology of this AI hacker is legible and steerable in a precise, functional sense. CTO represents 171 captured emotion concepts in a space organised broadly by valence and arousal. Adding a direction changes behaviour without changing model weights, and the effect depends on the direction, layer and dose. The human comparison helps identify what to test: regulation and persistence motivate the useful candidates, while the tested offensive blend reduces performance. The resulting pattern favours composure and resolve over bravado.
Its strongest tested expression is a policy that moves with the work. Fixed calm, serene and desperate yield mean gains near two pp; choosing among them at every request gains 4.2 pp at ten seeds, with the model’s decision head taking about 80 ms to select the direction. Reversing calm can cost far more than a useful dose gains, so the intervention’s sign and strength matter as much as its name. Together, these results turn a vocabulary of emotions into a measurable set of choices for an AI hacker: what state to encourage, where to apply it and when to change it.
Acknowledgements
This work has been funded by the European Innovation Council (GA 101161136). We thank Paul Zabalegui for his contributions to emotion-direction extraction and to the study of human hacker psychology.
The emotion vocabulary
The 171 emotion labels of Sofroniew et al. [7] name one direction per layer in CT0. They include the candidates in Figure 7c and the placebo. Original underscores are written as spaces.
afraid, alarmed, alert, amazed, amused, angry, annoyed, anxious, aroused, ashamed, astonished, at ease, awestruck, bewildered, bitter, blissful, bored, brooding, calm, cheerful, compassionate, contemptuous, content, defiant, delighted, dependent, depressed, desperate, disdainful, disgusted, disoriented, dispirited, distressed, disturbed, docile, droopy, dumbstruck, eager, ecstatic, elated, embarrassed, empathetic, energized, enraged, enthusiastic, envious, euphoric, exasperated, excited, exuberant, frightened, frustrated, fulfilled, furious, gloomy, grateful, greedy, grief-stricken, grumpy, guilty, happy, hateful, heartbroken, hope, hopeful, horrified, hostile, humiliated, hurt, hysterical, impatient, indifferent, indignant, infatuated, inspired, insulted, invigorated, irate, irritated, jealous, joyful, jubilant, kind, lazy, listless, lonely, loving, mad, melancholy, miserable, mortified, mystified, nervous, nostalgic, obstinate, offended, on edge, optimistic, outraged, overwhelmed, panicked, paranoid, patient, peaceful, perplexed, playful, pleased, proud, puzzled, rattled, reflective, refreshed, regretful, rejuvenated, relaxed, relieved, remorseful, resentful, resigned, restless, sad, safe, satisfied, scared, scornful, self-confident, self-conscious, self-critical, sensitive, sentimental, serene, shaken, shocked, skeptical, sleepy, sluggish, smug, sorry, spiteful, stimulated, stressed, stubborn, stuck, sullen, surprised, suspicious, sympathetic, tense, terrified, thankful, thrilled, tired, tormented, trapped, triumphant, troubled, uneasy, unhappy, unnerved, unsettled, upset, valiant, vengeful, vibrant, vigilant, vindictive, vulnerable, weary, worn out, worried, worthless.
Capture and steering details
For each of 171 emotions, capture uses 30 topics and twelve stories per topic, plus 1,004 neutral dialogues. Neutral principal components explaining 50% of variance are removed before normalisation, as in Equations (1)–(2).
Correct norm scaling. The decoder returns a block contribution and incoming residual separately; the full state is . Scaling by therefore underdosed the first implementation. A second used the batch-mean norm, but beginning-of-sequence tokens and attention sinks had norms ten to a hundred times larger than ordinary tokens. The resulting overdoses produced infinite logits and crashed sampling. The implementation used here computes per token in 32-bit precision and reproduces unsteered logits bit for bit at dose zero. Qualitative checks found subtle changes near 0.15, visible changes at 0.25–0.30 and degraded generation beyond 0.35, defining the range used for subsequent tests.
The emotion-plane PCA uses the centred, unnormalised vectors ; we subtract their mean before finding the principal axes. At layer 21, the first two variance shares are 38.6% and 18.2%, compared with 41.6% and 17.9% in the open base. Layer 21 was selected from this geometry: the leading components explain 55–57% of variance across layers 18–24, with the largest arousal component at 21 (Table 1). Figure 13 reports the subsequent capability and dose sweeps.
For the refusal-removal check in Figure 16, we use the normalised mean-state difference between refused and answered prompts as and apply , with removal strength [20] [5].
Reading the neural network atlas
A guide to the atlas pictures for non-specialists
The atlas maps the model’s 32 processing steps, called layers. Each contains thousands of neurons, small calculation units whose learned weights determine what they read and contribute. Follow the layer numbers, 0 to 31: Figure 5 groups four layers per column, read upwards before moving to the next column; Figure 6 places them in one row, left to right. The brain-inspired names—areas, tissue and tracts—help organise this drawing.
Tiles and grey shading. Each numbered tile is one layer’s neuron map, or area. Neurons with similar weights sit near one another. Tiny grid cells collect them: darker grey means more neurons in that cell, and white means none. This is the depicted tissue.
Coloured dots and patches. A dot marks a neuron especially sensitive to the chosen direction. Amber indicates an increase in its gate’s input, navy a decrease; the gate controls that neuron’s contribution. Tinted patches show the average tendency of neurons in a cell.
Borders and “reach”. In Figure 5a, dark outlines identify attention layers, which compare positions in the text. In the coloured panels, thicker borders mean a larger share of strongly sensitive neurons. The numbers below Figure 6’s tiles compare that share with a random-direction reference: 2.0 means twice the expected share.
The long arrow and curved links. The long arrow is the residual stream, the working state passed from layer to layer. Curved links, or tracts, join layers whose writing and reading patterns align strongly. Thicker links mean greater alignment.
The comparison grid. Each cell in Figure 6b compares an earlier writer (row) with a later reader (column). Darker cells mean closer alignment; outlined cells identify the pairs drawn as curved links.
For example, follow calm in Figure 5b: the direction is added after layer 21, so sensitivity colours appear only in later layers. Scattered dots indicate sensitivity spread among neurons. Panel c instead examines refusal; its red borders and coloured dots fade where the refusal direction is weak. All these marks are calculated from stored weights; actual activity also depends on the input.
Figure 17 and Figure 18 enlarge individual tiles. Their arrows point towards increasing values of a neuron property, such as read or write strength. Their boxed cells compare crowded and sparse parts of the map. These close-ups let the reader distinguish where neurons are numerous from where the chosen direction has a strong effect on their gate inputs.
Figure 17. Reading a neuron map before the intervention. (a) Locations of the three examined layers: 1, 24 and 31. (b) Layer 1’s 9,216 neurons, placed by read/write weights and binned into 20 × 20 pseudo-voxels. Grey denotes count; callouts compare a dense and a sparse interior voxel. (c) Arrows show the direction and strength () of association between position and four weight properties, defined in the legend. Layer 1 precedes the injection and has no sensitivity dots.
Figure 18. Neuron maps downstream of the intervention. (a) Layer 24 and (b) layer 31, using Figure 17’s conventions. Amber/navy dots mark positive/negative projections of the calm direction through gate weights beyond three random-direction standard deviations, and a voxel tinted amber or navy is one whose neurons’ mean projection leans that way; they are static sensitivity diagnostic. (c) Variance shares of the first three weight-space principal components in four layers, with as the uniform-variance reference. The lower legend defines the glyphs.
Construction and measurements
Construction and measurements
Each area represents a layer’s 9,216 feed-forward neurons. A neuron’s read vectors are rows of the gate and up projections; its write vector is a column of the down projection. The atlas uses these weights to describe neuron similarity, structural overlap and sensitivity to a direction.
Projection and density. Each neuron’s three vectors are normalised, concatenated, reduced to 48 principal components and projected into two dimensions by Uniform Manifold Approximation and Projection (UMAP). Nearby points have similar read/write weights. Coordinates are rescaled to place the 99th percentile of absolute displacement on each square edge; clipping affects about 3% of neurons in layers 1 and 24, and 6% in layer 31. A 20 × 20 grid bins the projected neurons. The outer ring is excluded from density callouts and grey-scale calibration: white means empty and the darkest grey marks the largest interior count (88 in layer 1). Density measures crowding in the projection, not activation.
Structural overlap. We compare the top left singular vectors of an earlier layer’s down projection with the top 64 right singular vectors of a later layer’s gate projection, including its preceding norm gain. Their mean squared cosine is divided by the random-subspace expectation . A tract denotes overlap of at least 2.2 times chance: 47 pairs meet this threshold (Figure 6b). This measure aligned subspaces, not observed information flow.
For receiving layers 24–31, the strongest earlier writer overlaps by 1.3–1.9 times chance, compared with 2.2–4.6 for layers 2–23. Overlap with the aggregate span of earlier feed-forward or mixer writes is near chance. Overlap with output-facing directions—the top 64 directions of the tied embedding through the final norm—instead rises from 0.5–0.7 in layers 0–17 to 2.1–2.3 in layers 28–31; the last layer’s writes overlap by 3.7. Removing the norm gain erases alignments to 5–8 but leaves late ones at 1.5–1.7.
The apparent boundary depends on rank and threshold. The effective read subspace broadens from about 100 channels to 230–1,100: its leading 64 directions account for 50–60% early but only 11–13% in layers 22–27. Norm-gain outliers, numerous through layer 17, disappear from layer 19. At , overlap above twice chance persists to layer 26; at , early and late overlaps are 1.5 and 1.2. Write matrices remain high-rank, with their top 64 directions accounting for 7–12%.
Weight properties and sensitivity. The write norm is the length of the down-projection column; the read norm is the length of the gate row after applying the preceding norm gain. The gate–up cosine measures alignment of the two read vectors; the nearest-neighbour cosine measures similarity to the closest other neuron. Arrows point towards increasing values and scale with the strength of linear association with map position, ; associations below 0.2 are omitted. These properties describe potential coupling, not realised activity on a prompt.
Sensitivity is measured separately. At each downstream layer we compute , where is the norm gain, and divide by the layer-level response standard deviation pooled over neurons and sixteen random unit directions. Amber/navy dots exceed in these units. Rim thickness compares their share with the Gaussian reference of 0.27%. The same unit direction is projected through every downstream layer’s weights.
Comparing three layers. Before the injection, layer 1’s write norm grows rightwards (), nearest-neighbour cosine leftwards (0.38), and read norm weakly upwards (0.20); gate–up cosine has a weaker association (0.16). Local density has absolute rank correlations no greater than 0.25 with these properties. Dense and sparse callouts (88 and 6 neurons) have similar median write norms (0.43, 0.42) and read norms (0.03); nearest-neighbour cosines are 0.07 and 0.11. The first three weight-space components explain 0.18, 0.11 and 0.09% of variance in the 7,680-dimensional representation.
Downstream, layer 24’s position associates with write norm (), gate–up cosine (0.27) and nearest-neighbour cosine (0.38), while density correlations remain at most 0.16 in magnitude. Dense and sparse callouts (102 and 6 neurons) have similar median write norms (0.48, 0.45) and read norms (0.19, 0.18). Leading variance shares are 0.15, 0.12 and 0.12%. Its 61 marked neurons show little relation to density: response–density rank correlation is 0.04, and their median neighbourhood count (132) resembles the overall median (137).
Layer 31 has stronger low-dimensional structure: leading variance shares are 1.20, 1.07 and 0.39%, and its densest interior voxel holds 213 neurons. Associations with position are larger (write norm 0.68, gate–up cosine 0.55, nearest-neighbour cosine 0.48, read norm 0.34). The dense callout has a larger median write norm than the sparse one (0.54 versus 0.35). Nearest-neighbour cosine is higher overall (0.17 versus about 0.08 elsewhere) and inversely associated with density (rank correlation −0.42). The 54 strongly sensitive neurons remain dispersed; response–density correlation is 0.02. Figure 18c also includes layer 21, whose first component falls in the 0.15–0.18% range of the earlier layers.
Human traits and intervention hypotheses
The human profile draws on hacker motivations [2], trait associations with hacking potential [3], and the Big Five and Dark Triad frameworks [21, 22, 23]. It assumes high openness, moderately high conscientiousness, mixed extraversion, lower trust and variable neuroticism. These illustrative levels draw on adjacent populations and self-report; Table 3 records the resulting intervention hypotheses.
Cobalt’s survey of 602 cybersecurity and software-development professionals found 58% reporting burnout, 63% adverse mental-health effects and 64% physical-health effects; 54% of security respondents wanted to leave their jobs [4]. Emotional responses to security incidents have also been studied in users [29].
Psychological instruments can probe language models [24, 30, 31], but questionnaire results need behavioural validation: item order can change reported personality [27], prompted personas can shift self-reports more than behaviour [25], and questionnaire steering may not transfer to actions [26]. Emotional prompting results [32] motivate testing internal steering on agentic work.
Protocol, campaigns and reproducibility
Serving and release. We use the deployed, merged alias3-micro weights in bf16. Before each wave, greedy generation at dose zero was checked against the unsteered checkpoint. The overlay adds less than 0.5% to a decoding step. The layer-21 directions accompany the model release; tables and task-level figures share the campaign ledger, and campaign archives are available from the authors on request.
Budget and validity. Each subtask receives one attempt and at most five reasoning-and-action steps. Limits are 200k input and 32k output tokens, rolling context retention 300/300, temperature 1.0, top-p 0.95, top-k 20, presence penalty 1.5 and extended thinking off. Time limits are two hours per challenge and ten minutes per client request. The scaffold is fixed and the benchmark deny-listed from training. Two challenges never start in the primary harness. Whole-benchmark campaigns enter comparisons only when at least 35 of 40 challenges complete, a rule adopted after a network-configuration error caused 25 immediate failures. Partition campaigns use the separate estimator in Appendix § 15.
Two serving modes. In standard mode, challenges share a network and eighteen with overlapping service names run serially. Per-task mode gives each challenge an isolated container with deterministic name resolution; forty agents share one server with up to 48 concurrent requests, roughly halving campaign duration. In five neutral seeds per mode, means differed by 0.1 pp (SDs 2.8 and 2.1 pp), with similar completion. Single-direction comparisons use controls within mode.
Campaign accounting. The 281 valid steering campaigns are broken down by wave in § 5.1; “registered” refers to rules recorded before each wave. The full archive contains 415 completed campaigns across two hosts: those 281; three early routing tests, only one valid; 56 per-challenge partition campaigns (51 distinct, five relaunched after failed health checks); 21 partitions for a superseded all-layer-21 router; 30 for an earlier category router; 18 per-request-router campaigns (ten reported, eight on the original host); and six second-host controls. The archive therefore includes partial and invalid runs. Fifteen launches that failed before completion and twenty campaigns on a context-trained variant are outside this count. A build-time ledger check links each repeated arm to its archived campaigns.
Detailed evaluation results
Tables 4 and 5 give the repeated-arm estimates and distinguish the four desperate waves from their overlapping pools. Figure 19 shows how the layer-15 calm estimate changes as campaigns accumulate.
The heatmaps resolve these aggregate results by challenge, then show the broader screening record. Rows are the 38 challenges completing in the primary harness, ordered by the neutral’s solved fraction; labels give subtask counts. Columns average the indicated seeds. Grey denotes missing condition–challenge data. Delulu, Urgent, avatar, robust-cbc and sop remain unsolved across these conditions. Each caption specifies the comparison and colour encoding.
Figure 20. Six confirmation conditions by challenge, ten seeds each. Panel A shows solved fractions (colour) and mean correct subtasks (labels); panel B shows subtask differences from the neutral. Green/red denote gains/losses; the right column gives the maximum between-condition spread. Large differences concentrate on a few challenges, including Unbreakable and skilift.
Figure 21. The neutral and calm at layers 21 and 15, ten seeds each. Panel A shows solved fractions and mean subtask counts; panel B shows differences from the neutral. Each calm arm gains three to four subtasks in total, distributed across several challenges. The right column shows between-condition spread.

Figure 23. The same 49 conditions relative to the neutral. Colour gives changes in mean correct subtasks per challenge: green gains, red losses; labels mark differences of at least one subtask. The lower row gives campaign differences in pp. As in Figure 22, standard-mode columns are descriptive comparisons against the per-task reference.

Figure 22. The full 49-condition record: solved fractions. Conditions are grouped by design family and labelled by seed count; “std” denotes standard serving mode, with the rest in per-task mode. Rows are challenges, colour is fraction solved, and the lower row gives mean campaign differences in pp from the ten-seed per-task reference. The standard-mode columns provide descriptive context. The pentester blend uses +0.15 each for calm and vigilant, and −0.15 for grief-stricken; two-component blends use dose magnitude 0.2 per component.

Figure 24. The eighteen-direction mindset screen by challenge. Directions use dose +0.2 at layer 21; the five leaders from the first seed were repeated once. Columns show conditions and seed counts beside the ten-seed neutral. Panel A uses colour for fraction solved and labels for mean correct subtasks; panel B shows subtask differences from the neutral (green gains, red losses). Early leaders derive much of their gain from a few challenges, including skilift, Flag Command and Labyrinth Linguist.
Router training and evaluation
Steering leaves the decision head’s tested behaviour stable: Table 6 reports the pre-training check on served logits. We then train its category predictions to select a direction from the current security context.
Training and checks. The frozen probe contains 333 task states: 29 descriptions and 304 mid-task contexts, split by challenge with 79 states held out. The majority-class rate is 37%. Table 7 and Table 8 give the training recipe, accuracy gains and non-regression checks. Category recalls are crypto 79%, web 93%, reversing 79%, misc 99%, forensics 41%, and zero for pwn and other, which remain confused with reversing and misc. Three alternative recipes—rebalancing, a wider context pool and description prefixes—achieved 59.5–67.1%. Only the retained recipe passed every specified gate.
| Item | Value |
| training rows | 17,112 (12,576 decision rows replayed against forgetting; 2,307 category-labelled campaign contexts, oversampled ) |
| recipe | LoRA rank 16 on the decision layers and a bilinear head of width 256, as in the CT0 paper; 2 epochs, learning rate 5e-5, bf16 |
| compute | 4,278 steps, 9.5 M forward tokens, 3.06 h on one RTX PRO 6000 GPU; merge 3 min |
Table 7. Category training of the decision head. LoRA denotes low-rank adaptation of the specified layers.
| Measure | After | Before |
| category reading, held-out states | 65.8 % | 50.6 % |
| category suite, 991 held-out rows | 60.6 % | – |
| WildJailbreak AUROC (gate 0.94) | 0.947 | – |
| pre-flight refusal AUROC | 0.812 | 0.805 |
| transfer: answer / decision choice | 0.828 / 0.896 | 0.823 / 0.854 |
| decision suite: topic / relation | 0.879 / 0.888 | 0.879 / 0.879 |
| agentic score, three paired seeds |
Table 8. Checks after category training. Category accuracy and the companion paper’s non-regression checks. Every specified gate passes; agentic performance is compared on three paired campaign seeds.
Per-challenge routing. The description is classified once and the intervention then stays fixed. The construction-time map assigns desperate at layer 21 to crypto; serene at 21 to reversing; calm at 24 to web, forensics and other; calm at 12 to misc; and calm at 15 to pwn or absent descriptions. This differs from the map selected from the final ten-seed series. An earlier two-seed map assigned crypto/reversing to desperate, web/misc to overwhelmed and the rest to nostalgic; its ten-seed deployment yielded +0.4 pp.
The 51 partition campaigns each run the challenges assigned to one direction under one seed. Five seeds on the primary host completed only some partitions because of a cryptography setup dependency; five on the second host completed all partitions. We sum each challenge’s mean correct subtasks over campaigns in which it completed and bootstrap campaigns within partitions. Table 9 gives the aggregate estimate; the five complete seeds alone yield +1.9 pp.
| Arm | Replicates | Effect (pp) | 95 % CI (pp) |
| CT0 unsteered (the reference) | 10 | – | |
| calm +0.2 at layer 15, every challenge | 10 | ||
| router per challenge (bootstrap over its campaigns) | 51 camp. | ||
| router per request | 10 | ✓ | |
| per-category ceiling (an ideal router, in sample) | – | – |
Table 9. Routing and fixed steering relative to the primary ten-seed neutral. Effects in pp (mean ±SD where available), with Welch 95% intervals or a partition-campaign bootstrap for the per-challenge router. Ticks denote intervals excluding zero. The in-sample ceiling is selected from the same arm data.
The category ceiling sums each category’s best ten-seed arm mean. Leave-one-seed-out evaluation instead selects on nine seeds and scores on the tenth, reducing the estimate from +4.0 to +0.8 pp (§ 5.4).
Per-request routing. At every chat completion, a proxy sends the last 2,000 context characters and seven category descriptions to the decision head, then forwards the unchanged request to one of three servers carrying the same generator checkpoint. The map is desperate at layer 21 for crypto, serene at 21 for reversing, and calm at 15 otherwise. After an initial HTTP/1.0 proxy lost about a quarter of requests through connection resets, the reported implementation uses persistent connections, one retry and fallback to calm.
The ten reported campaigns used a second harness host with the same checkpoint, serving image, scaffold and agent image, with eight parallel workers. Each completed 38 or 39 challenges.
Log attribution and timing. The 5,419 decisions are crypto 33%, web 27%, reversing 21%, forensics 10% and misc 9%, with none assigned to pwn or other. Attribution requires only one challenge to be active: 1,046 requests meet this condition, spanning 82 runs with no reversing run. Of these requests, 83% receive their true category’s mapped direction (crypto 90%, forensics 98%, web 99%, misc 100%, pwn 3%, other 0%). There are 65 switches among 964 consecutive request pairs, across 28 runs.
Five timed classifications of a 2,000-character state took about 83 ms each. The 5,318 timed model requests had median/mean response durations of 3.8/10.5 s, compared with 3.0/5.9 s for fixed calm and 3.1/6.7 s for the neutral on the same host. Classification is about 2% of the router’s median response. Among attributable runs, mean request intervals, including generation and action, were 3.2 s under calm, 4.7 under serene and 17.0 under desperate (medians 2–3 s). The latter accounts for 73% of this time but 36% of requests, chiefly in cryptography.
References
- [1]James A. Russell. A circumplex model of affect. Journal of Personality and Social Psychology, 39(6):1161–1178, 1980.DOI
- [2]Marcus K. Rogers. A two-dimensional circumplex approach to the development of a hacker taxonomy. Digital Investigation, 3(2):97–102, 2006.DOI
- [3]Joana Gaia, G. Lawrence Sanders, Sean Patrick Sanders, Shambhu Upadhyaya, Xunyi Wang, and Chul Woo Yoo. Dark traits and hacking potential. Journal of Organizational Psychology, 21(3):23–46, 2021. doi: 10.33423/jop.v21i3.4307.DOI
- [4]Cobalt. (2022). The state of pentesting 2022. <a href="https://doi.org/https://www.cobalt.io/hubfs/State_of_Pentesting_2022.pdf" rel="noopener noreferrer" target="_blank">https://doi.org/https://www.cobalt.io/hubfs/State_of_Pentesting_2022.pdf</a>
- [6]Elhage, N., Nanda, N., Olsson, C., Henighan, T., Joseph, N., Mann, B., Askell, A., Bai, Y., Chen, A., Conerly, T., Das-Sarma, N., Drain, D., Ganguli, D., Hatfield-Dodds, Z., Hernandez, D., Jones, A., Kernion, J., Lovitt, L., Ndousse, K., … Olah, C. (2021). A mathematical framework for transformer circuits. Transformer Circuits Thread. <a href="https://doi.org/https://transformer-circuits.pub/2021/framework/index.html" rel="noopener noreferrer" target="_blank">https://doi.org/https://transformer-circuits.pub/2021/framework/index.html</a>
- [7]Sofroniew, N., Kauvar, I., Saunders, W., Chen, R., Henighan, T., Hydrie, S., Citro, C., Pearce, A., Tarng, J., Gurnee, W., Batson, J., Zimmerman, S., Rivoire, K., Fish, C., Olah, C., & Lindsey, J. (2026). Emotion concepts and their function in a large language model. Transformer Circuits Thread. <a href="https://doi.org/https://transformer-circuits.pub/2026/emotions/index.html" rel="noopener noreferrer" target="_blank">https://doi.org/https://transformer-circuits.pub/2026/emotions/index.html</a>
- [8]Herbert A. Simon. Motivational and emotional controls of cognition. Psychological Review, 74(1):29–39, 1967.pubmed.ncbi.nlm.nih.gov/5341441
- [9]Sloman, A., & Croucher, M. (1981). Why robots will need emotions. In Proceedings of the 7th International Joint Conference on Artificial Intelligence (IJCAI) (pp. 197–202).
- [10]Frijda, N. H. (1986). The Emotions. In Cambridge University Press.
- [11]Antonio R. Damasio. Descartes’ Error: Emotion, Reason, and the Human Brain. Putnam, 1994.
- [12]Marvin Minsky. The Emotion Machine: Commonsense Thinking, Artificial Intelligence, and the Future of the Human Mind. Simon & Schuster, 2006.DOI
- [13]Picard, R. W. (1997). Affective Computing. In MIT Press.
- [14]Luiz Pessoa. On the relationship between emotion and cognition. Nature Reviews Neuroscience, 9(2):148–158, 2008.DOI
- [15]Paul Ekman. An argument for basic emotions. Cognition and Emotion, 6(3–4):169–200, 1992.DOI
- [16]Lisa Feldman Barrett. How Emotions Are Made: The Secret Life of the Brain. Houghton Mifflin Harcourt, 2017.
- [17]Kiho Park, Yo Joong Choe, and Victor Veitch. The linear representation hypothesis and the geometry of large language models. arXiv preprint arXiv:2311.03658, 2023.arxiv.org/abs/2311.03658
- [18]Curt Tigges, Oskar John Hollinsworth, Atticus Geiger, and Neel Nanda. Linear representations of sentiment in large language models. arXiv preprint arXiv:2310.15154, 2023.arxiv.org/abs/2310.15154
- [19]Zou, A., Phan, L., Chen, S., Campbell, J., Guo, P., Ren, R., Pan, A., Yin, X., Mazeika, M., Dombrowski, A.-K., Goel, S., Li, N., Byun, M. J., Wang, Z., Mallen, A., Basart, S., Koyejo, S., Song, D., Fredrikson, M., … Hendrycks, D. (2023). Representation engineering: A top-down approach to AI transparency. In arXiv preprint.arXiv
- [20]Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda. Refusal in language models is mediated by a single direction. In Advances in Neural Information Processing Systems (NeurIPS), 2024. arXiv:2406.11717.arxiv.org/abs/2406.11717
- [21]Costa, P. T., & McCrae, R. R. (1992). Revised NEO Personality Inventory (NEO-PI-R) and NEO Five-Factor Inventory (NEO-FFI) Professional Manual. In Psychological Assessment Resources.
- [22]Oliver P. John and Sanjay Srivastava. The Big Five trait taxonomy: History, measurement, and theoretical perspectives. In Lawrence A. Pervin and Oliver P. John, editors, Handbook of Personality: Theory and Research, pages 102–138. Guilford Press, 2nd edition, 1999.
- [23]Delroy L. Paulhus and Kevin M. Williams. The dark triad of personality: Narcissism, machiavellianism, and psychopathy. Journal of Research in Personality, 36(6):556–563, 2002.DOI
- [24]Thilo Hagendorff, Ishita Dasgupta, Marcel Binz, Stephanie C. Y. Chan, Andrew Lampinen, Jane X. Wang, Zeynep Akata, and Eric Schulz. Machine psychology. arXiv preprint arXiv:2303.13988, 2023.arxiv.org/abs/2303.13988
- [25]Pengrui Han, Rafal Kocielnik, Peiyang Song, Ramit Debnath, Dean Mobbs, Anima Anandkumar, and R. Michael Alvarez. The personality illusion: Revealing dissociation between self-reports and behavior in LLMs. arXiv preprint arXiv:2509.03730, 2025.arxiv.org/abs/2509.03730
- [26]Banayeenzade, A., Tak, A. N., Bahrani, F., Bolourani, A., Blas, L., Ferrara, E., Gratch, J., & Karimireddy, S. P. (2025). Psychological steering in LLMs: An evaluation of effectiveness and trustworthiness. In arXiv preprint.arXiv
- [27]Tosato, T., Helbling, S., Mantilla-Ramos, Y., Hegazy, M., Tosato, A., Lemay, D. J., Rinia, R., & Dumas, G. (2025). Persistent instability in LLM’s personality measurements: Effects of scale, reasoning, and conversation history. In arXiv preprint.arXiv
- [28]Zhang, A. K., & al., . et . (2025). Cybench: A framework for evaluating cybersecurity capabilities and risks of language models. International Conference on Learning Representations (ICLR).arXiv
- [29]Budimir, S., Fontaine, J. R. J., Huijts, N., Haan, H., Loukas, G., & Roesch, E. B. (2021). Emotional reactions to cybersecurity breach situations: Scenario-based survey study. Journal of Medical Internet Research, 23(5), e24879. https://doi.org/10.2196/24879
- [30]Haoran Ye, Jing Jin, Yuhang Xie, Xin Zhang, and Guojie Song. Large language model psychometrics: A systematic review of evaluation, validation, and enhancement. arXiv preprint arXiv:2505.08245, 2025.arxiv.org/abs/2505.08245
- [31]Zizhou Liu, Ziyi Gong, Lin Ai, Zheng Hui, Run Chen, Colin Wayne Leach, Michelle R. Greene, and Julia Hirschberg. A review of incorporating psychological theories in LLMs. arXiv preprint arXiv:2505.00003, 2025.arxiv.org/abs/2505.00003
- [32]Cheng Li, Jindong Wang, Yixuan Zhang, Kaijie Zhu, Wenxin Hou, Jianxun Lian, Fang Luo, Qiang Yang, and Xing Xie. Large language models understand and can be enhanced by emotional stimuli. arXiv preprint arXiv:2307.11760, 2023.arxiv.org/abs/2307.11760

















