Security flaw in major AI provider APIs exposed hidden reasoning and user secrets
Researchers discovered that encrypted reasoning objects in APIs from major AI providers could be decoded to recover hidden information including API keys and passwords from user sessions.
Researchers have found a flaw in how OpenAI, Anthropic, and Google handle encrypted reasoning objects in their AI APIs, according to The Hacker News. The weakness allowed researchers to recover hidden reasoning and secrets from session logs, including 62 API keys, 33 passwords, 24 access tokens, and seven private keys recovered from 704 user sessions. Weaker models in the same provider family could be used to decode reasoning blocks created by stronger models, and opaque reasoning blocks could be replayed across different sessions and users.
The attack required obtaining an encrypted reasoning block and API access to a compatible model from the same provider. The researchers disclosed findings to the affected providers, and say the demonstrated attacks stopped working after mitigations by August 2026. Developers are advised to strip reasoning blocks from shared traces and avoid committing raw API transcripts, particularly when conversation state is managed manually or statelessly. Neither OpenAI, Anthropic, nor Google has publicly acknowledged the flaw.