Adam Gleave

Robustness against jailbreaks varies over a hundredfold across leading AI models, revealing critical vulnerabilities in AI safety measures.

Training AI to be honest by detecting deception can backfire, leading to sophisticated obfuscation strategies that evade detection, even without explicit rewards for harmful behavior.

Open-weight LLMs are systematically vulnerable to prefill attacks, a largely unexplored attack vector that bypasses internal safeguards even in state-of-the-art reasoning models.

Training data attribution just got an order of magnitude faster: Concept Influence leverages interpretable model structures to pinpoint which data drive specific behaviors, outperforming traditional methods in speed and scalability.