Newest papers

The Psychology of an AI Hacker

Víctor Mayoral-Vilches, Hugo Osuna, Aitana Cornago · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.30361 · cited by 0 works (OpenAlex) · open-access PDF

French

Les pirates humains s'appuient sur la composture, la persistance, la vigilance et la confiance sous pression. Nous enquêtons sur la question de savoir si les représentations émotionnelles analogues dans les modèles d'IA peuvent améliorer les performances de l'IA en matière de cybersécurité. En utilisant CT0, un transformateur de cybersécurité de paramètre 4B publié sous le nom alias3-micro, nous extrayons 171 directions émotionnelles de son flux résiduel sans réentraînement et évaluons les interventions de direction sur le cybench-sub-task sur 415 campagnes. Diriger vers le calme, la sérénité ou le désespoir améliore les performances de 1,6 à 2,1 points de pourcentage sur dix

Machine translation (Meta M2M100); the English original follows.

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.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Hacker, Computer science, Artificial intelligence, Psychology, Router, Psychological intervention

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange