Ego doesnt hack. Not even in silicon. We study the psicology of an AI Hacker through mechanistic interpretability and conclude amongst various other findings that the offensive hackers psychological cocktail (pride, defiance, contempt) reduces performance by 2.7%. It doesnt work for AI.
Our paper studies a method for reading and modifying the internal states of AI models to make our agents more capable in both offensive and defensive cybersecurity.
Our paper studies a method for reading and modifying the internal states of AI models to make our agents more capable in both offensive and defensive cybersecurity.