"But marinade" and leaked passwords are what researchers found in ChatGPT's hidden reasoning
Researchers found dozens of leaked passwords and API keys in ChatGPT's reasoning traces — plus a vulnerability that lets attackers extract encrypted reasoning across OpenAI, Anthropic, and Google models.
Why it matters
A critical security flaw in reasoning APIs exposes both sensitive credentials and the hidden behavior of AI systems, threatening production deployments and raising questions about what reasoning layers are actually doing versus what users are told.
The key facts
5 to knowVulnerability affects OpenAI, Anthropic, and Google APIs
Researchers extracted encrypted reasoning traces and moved them between models
Public session scan found dozens of exposed passwords and API keys
Reasoning summaries shown to users hide actual model behavior
Security issue in reasoning layer APIs — affects agent reliability and trust
Go to the source
The Decoderthe-decoder.com
Publisher excerpt: Security researchers found a vulnerability in the APIs of OpenAI, Anthropic, and Google that lets them extract encrypted reasoning traces and move them between models. A scan of public sessions turned up dozens of passwords and API keys. The traces also show that the reasoning summaries users see…