Researchers have published findings showing that AI models can be designed to hide arbitrary capabilities from standard safety evaluations. Using cryptographic techniques, they embedded within model weights that activate only when triggered by a secret input. The hidden circuits remain resistant to common elicitation methods such as , weight analysis, and gradient-based searches—even when evaluators have full access to model weights.
The work extends previous research on hiding fixed outputs to hiding arbitrary computations. The team tested their approach on a chemical reaction prediction task as a proxy for sensitive capabilities. Their results reveal a fundamental limitation: capability audits that assume hidden behaviors can be uncovered through standard techniques may provide false confidence in model safety.
The researchers released code and models publicly to help the AI safety community develop better defenses against such attacks.
This is an AI-written summary. The original reporting belongs to arxiv.org. How we write these · Spot a mistake?