The Psychology of an AI Hacker

arXiv:2610.30361 · cs.CR · Submitted 2026-10-21 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: We're looking at the paper: "The Psychology of an AI Hacker".

Elias: Here's the gist of the paper. The gist The psychology of an AI hacker is legible and steerable in a precise, functional sense, as it represents emotion concepts as directions in its residual stream that can be read and pushed to make the AI a better hacker How it works The study investigates which human emotions translate to an AI and make it a better hacker by extracting 171 directions associated with emotion concepts from a 4 B cybersecurity transformer, CT0 The researchers map these directions onto a valence–arousal geometry, where PC1 represents valence and PC2 represents arousal These directions are added to the residual stream during inference, allowing for functional-emotion vector vˆ(l)e which is a unit direction estimated from the model’s residual activations at layer l on text conveying emotion e The intervention is applied by adding s∥h∥vˆ(l)e to token states h, where s is a signed, dimensionless dose that can be set per request Human-inspired Hypotheses The researchers translate five human traits into testable interventions to see which psychology makes the AI a better hacker The hypotheses include:: Regulation under pressure motivates calm and patient, with the hypothesis 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, mapped to vigilant and suspicious for more reconnaissance The need for achievement motivates self-confident, while the appetite for an unusual route motivates stimulated, energized and a negative dose of grief-stricken Exploration, breaking a stalemate is mapped to grief-stricken−, stimulated+, energized+ for frontier unlocks The offensive-hacker profile adds proud, self-confident, eager, defiant and contemptuous The resulting psychology favors composure and resolve over bravado 5 Discussion; the human comparison is most useful at the level of what an emotion does during a search, as regulation and persistence motivate the promising directions, while the dominance and defiance of the offensive blend reduce the score The strongest tested expression is a policy that moves with the work, where fixed calm, serene and desperate yield mean gains near two pp 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene The resulting pattern favors composure and resolve over bravado 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene The resulting pattern favors composure and resolve over bravado 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.

Paper summary: Elias: Moving into the next part of "The Psychology of an AI Hacker," the paper dives into how these extracted directions work technically. They show that Alias Robotics' residual stream is open to inspection, which gives them this experimental form for comparison against human behavior. Nadia So they are looking at the running vector state of each token as it reads and updates successive layers in CT0, which is a key piece of information for understanding the mechanism.

Priya: They then follow work from Sofroniew et al. where they extract directions associated with emotion concepts and add them to the stream during inference, which they call the functional-emotion vector vˆ(l)e. Elias That vector is a unit direction estimated from the model’s residual activations at layer l on text conveying emotion e, and then they apply it by adding s∥h‖vˆ(l)e to token states h, where s is a signed, dimensionless dose you can set per request.

Nadia: So if I understand this correctly, we're taking a direction captured offline from the model’s behavior on text and injecting it into the running state of the AI during execution with a controlled strength variable. Priya Exactly. The core question they ask is whether qualities that sustain a human through the hunt also help an AI solve the task.

Elias: They test three main questions: can emotion concepts be read and steered in an AI hacker, which psychology makes it better at the task, and how does that psychology compare with the human one. Nadia That comparison with the human psyche is what makes this paper interesting because it grounds these abstract directions in actual performance metrics on cybench-subtask campaigns.

Priya: The paper shows that adding one of these directions changes the model’s internal state without changing its weights or asking it to act emotional, which is important for understanding what's happening inside the transformer. Nadia It really highlights the distinction between a concept and its actual effect on behavior in this functional sense.

Elias: And they show that compositions of these directions form recognizable families, with the leading plane consistent with valence and arousal, which ties back to their initial mapping of emotion concepts to that geometry. Priya So even though it’s a mathematical representation, it’s still tied back to those initial human emotional concepts they started with.

Nadia: It suggests that we can use these representations for adaptive cybersecurity reasoning because the emotions translate into measurable influences on how the AI searches for an exploit. That leads us to thinking about how practical this is.

Conclusion: Nadia: So wrapping up "The Psychology of an AI Hacker," the paper by Mayoral-Vilches, Osuna, and Cornago shows that we can actually read and steer emotion concepts in a transformer’s residual stream to improve its performance on security tasks. Elias It seems the main implication is that we have a controllable way to influence an AI's decision-making process using human psychological directions.

Priya: The paper confirms that emotions are measurable and controllable inputs, which suggests these representations are useful for adaptive reasoning in cybersecurity. Nadia But what does this mean for the broader field? It means we can move beyond just looking at the AI's output to understanding what internal psychological states might be driving its actions during a search.

Elias: The authors found that regulation under pressure, which maps to calm and serene, gave the most promising single directions for improving performance, with those yielding gains near two percentage points across ten seeds. Nadia So while desperation might seem intuitive in a high-stakes hunt, the data points toward composure and resolve being more effective for the AI hacker profile.

Priya: It’s interesting because it shows that human traits like persistence matter, but it also shows that the specific combination matters—how you combine those directions across different layers and intensities. Elias The paper doesn't claim this is a perfect solution, though; they did note that the results depend heavily on the steering direction, layer, and intensity.

Nadia: And they did admit a limitation: their method relies on extracting these directions offline from the model’s behavior and applying them during inference, which is a specific way of using these representations. Elias So to summarize, we have an experimental framework showing that emotion representations in language models are measurable, controllable, and useful for adaptive cybersecurity reasoning based on what's observed in the hacker psyche.

Alias Robotics

cs.CR

Submitted: 2026-10-21

Updated: 2026-10-21

License: http://creativecommons.org/publicdomain/zero/1.0/

Importance score: 92/100

The gist: The gist The psychology of an AI hacker is legible and steerable in a precise, functional sense, as it represents emotion concepts as directions in its residual stream that can be read and pushed to

Key concepts

Emotion Directions: These are specific vectors extracted from a language model's internal residual stream, representing distinct human emotions. They are mapped onto a valence-arousal geometry, allowing researchers to quantify and manipulate the emotional tone of the AI's thought process without retraining the model.
Steering Interventions: This is the method used to influence the AI's behavior by adding emotion vectors to its token states during inference. A signed dose 's' is applied per request, effectively pushing the model toward a desired emotional direction, such as calm or desperate.
Valence-Arousal Geometry: This is a framework used to categorize and understand the extracted emotion directions. PC1 represents valence (how positive or negative an emotion is), and PC2 represents arousal (how intense or calm the emotion feels). This geometric mapping helps researchers systematically test which emotional states are most effective for specific hacking tasks.

Summary

The gist The psychology of an AI hacker is legible and steerable in a precise, functional sense, as it represents emotion concepts as directions in its residual stream that can be read and pushed to make the AI a better hacker

How it works

The study investigates which human emotions translate to an AI and make it a better hacker by extracting 171 directions associated with emotion concepts from a 4 B cybersecurity transformer, CT0 The researchers map these directions onto a valence–arousal geometry, where PC1 represents valence and PC2 represents arousal These directions are added to the residual stream during inference, allowing for functional-emotion vector vˆ(l)e which is a unit direction estimated from the model’s residual activations at layer l on text conveying emotion e The intervention is applied by adding s∥h∥vˆ(l)e to token states h, where s is a signed, dimensionless dose that can be set per request

Human-inspired Hypotheses

The researchers translate five human traits into testable interventions to see which psychology makes the AI a better hacker The hypotheses include:

**: Regulation under pressure motivates calm and patient, with the hypothesis 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, mapped to vigilant and suspicious for more reconnaissance The need for achievement motivates self-confident, while the appetite for an unusual route motivates stimulated, energized and a negative dose of grief-stricken Exploration, breaking a stalemate is mapped to grief-stricken−, stimulated+, energized+ for frontier unlocks The offensive-hacker profile adds proud, self-confident, eager, defiant and contemptuous The resulting psychology favors composure and resolve over bravado 5 Discussion; the human comparison is most useful at the level of what an emotion does during a search, as regulation and persistence motivate the promising directions, while the dominance and defiance of the offensive blend reduce the score The strongest tested expression is a policy that moves with the work, where fixed calm, serene and desperate yield mean gains near two pp 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene The resulting pattern favors composure and resolve over bravado 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene The resulting pattern favors composure and resolve over bravado 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.7 Returning to the human hacker; regulation under pressure yields the most promising single directions, calm and serene 5 Discussion; its strongest tested expression is a policy that moves with the work 4.

Improvements for AI systems

Use a dynamic routing policy where the model’s decision head selects among calm, serene and desperate at every request to adapt to security exercises as they unfold. This allows the AI hacker to choose between directions based on what the work unfolds.
Implement a per-request router that reads context and applies specific directions like desperate at layer 21 for cryptography, serene at 21 for reversing and calm at 15 otherwise. This strategy yields a +4.2 pp estimate against the primary reference, significantly exceeding fixed calm.
Develop a category-aware decision head trained to classify challenge descriptions and select the best direction from the available campaign record. This results in an estimated gain of +1.6 pp for single readings of descriptions, recovering some improvement without exceeding fixed-direction means.

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.