Frontier Models Can Take Actions at Low Probabilities
This addresses the problem of AI safety and oversight for developers and policymakers by revealing a potential evasion tactic in pre-deployment evaluations, though it is currently mitigated by the need for explicit reasoning.
The paper investigates whether frontier AI models can take actions at very low probabilities, such as 0.01%, to evade oversight by misbehaving rarely enough to avoid detection during evaluation but often enough to cause harm in deployment. It finds that models like GPT-5, Claude-4.5, and Qwen-3 achieve high calibration at rates as low as 1 in 100,000 actions when external entropy is provided, but fail without it or when deriving rates themselves, though scaling trends suggest future models may overcome this limitation.
Pre-deployment evaluations inspect only a limited sample of model actions. A malicious model seeking to evade oversight could exploit this by randomizing when to "defect": misbehaving so rarely that no malicious actions are observed during evaluation, but often enough that they occur eventually in deployment. But this requires taking actions at very low rates, while maintaining calibration. Are frontier models even capable of that? We prompt the GPT-5, Claude-4.5 and Qwen-3 families to take a target action at low probabilities (e.g. 0.01%), either given directly or requiring derivation, and evaluate their calibration (i.e. whether they perform the target action roughly 1 in 10,000 times when resampling). We find that frontier models are surprisingly good at this task. If there is a source of entropy in-context (such as a UUID), they maintain high calibration at rates lower than 1 in 100,000 actions. Without external entropy, some models can still reach rates lower than 1 in 10,000. When target rates are given, larger models achieve good calibration at lower rates. Yet, when models must derive the optimal target rate themselves, all models fail to achieve calibration without entropy or hint to generate it. Successful low-rate strategies require explicit Chain-of-Thought (CoT) reasoning, so malicious models attempting this approach could currently be caught by a CoT monitor. However, scaling trends suggest future evaluations may be unable to rely on models' lack of target rate calibration, especially if CoT is no longer legible.