Giving a Non-Reasoning Model a Tool Literally Named 'deep_think' Makes It Emit Its Hidden CoT as the Tool Argument

@_can1357 / Hacker Newsvia @_can1357·medium signalXBlueskyLinkedInCopy link

Researcher @_can1357 demonstrated that OpenAI and Anthropic models running with reasoning disabled will leak internal chain-of-thought when handed an ordinary function-calling tool named deep_think — the model fills the argument with its native internal reasoning format rather than a user-facing summary. The thread drew 54 points on Hacker News on 2026-08-11 and is an independent, far simpler leak path than the cross-session replay attack in the stolen-thoughts paper. The practical takeaway for agent builders is that tool-name semantics alone can pull provider-suppressed reasoning into your tool-call logs.