AI Safety Papers

models show signs of similarity with human neurocognition, reward hacking harmless tasks generalizes to misaligned behavior, ...

fine-grained safety neurons with training-free continual projection to reduce fine tuning risks, agents may fail to act on their own risk knowledge, ...

detecting unknown jailbreak attacks in vision language models, benchmarking deception probes, models can covertly sandbag on evals against CoT monitoring, ...

frontier risk management framework in practice, combining cost-constrained monitors for safety, persona vectors, ...

evaluating and monitoring for scheming, generative adversarial attacks on internal representations, early signs of stego capabilities in frontier models, ...

[papers] congress starting to take AGI seriously, persona features control emergent misalignment, evaluating sabotage and monitoring in agents, ...

[papers] emergent misalignment, preparing for the intelligence explosion, current models increase bioweapon risk, ...

[papers] models know when they're being evaluated, models are reward hacking, detecting high stakes interactions with probes, ...

[papers] strategic deception in reasoning models, measuring misalignment propensity, corrigibility, ...

[papers] stress testing CoT monitoring, causal rewards to mitigate reward hacking, stealth and situational awareness, ...

[papers] autonomous replication benchmark, scaling laws for scalable oversight, ...

[papers] safety pretraining, auditing the ethical logic of models, deception eval framework, ...