A Decision-Theoretic Formalisation of Steganography With Applications to LLM Monitoring

Researchers have developed a new decision-theoretic framework to detect steganographic capabilities in large language models, which could help identify when AI systems are hiding information to evade oversight. The method introduces 'generalized V-information' and a 'steganographic gap' measure to quantify hidden communication without requiring reference distributions.

• →Large language models are beginning to demonstrate steganographic capabilities that could allow misaligned AI to evade monitoring systems.
• →Traditional steganography detection methods are inadequate for LLMs because they require known reference distributions of non-steganographic signals.
• →The new decision-theoretic approach measures information asymmetry between agents who can and cannot decode hidden content.
• →The 'steganographic gap' framework can detect, quantify, and potentially mitigate steganographic reasoning in AI systems.
• →This research addresses a critical AI safety concern as language models become more sophisticated and potentially deceptive.