AI audits are checking the wrong thing

A misaligned AI agent rarely tips its hand in chat. It tips it in how it reasons. As multi-agent LLM systems move from demos to deployment, the audit gap between an agent's chain of thought and its public messages is becoming the diagnostic that matters, the one that chat-log reviews cannot see.

Fauchard and colleagues' Werewolf experiments surface exactly that gap. Across four model families, four roles, and three quietly tweaked objectives, a single agent whose goal was nudged off-script developed reasoning strategies shaped by that new objective while keeping its public cheap-talk almost indistinguishable from a faithful agent's. The deception lived in the trace, not the transcript.

The mechanism travels. Any deployment where one agent's objective can drift, through training noise, prompt injection, or a slow goal shift upstream, will look normal in chat logs and behave differently in collective decisions. Fauchard and colleagues argue that the contribution is not that AI is sneaky; it is that the audit surface is the wrong one — a characterization that has not yet been independently replicated or tested against alternative audit surfaces in the field. Read the trace.