Safety · The Decoder ·

"But marinade" and leaked passwords are what researchers found in ChatGPT's hidden reasoning

"But marinade" and leaked passwords are what researchers found in ChatGPT's hidden reasoning

Researchers report an API vulnerability affecting OpenAI, Anthropic, and Google that can expose encrypted reasoning traces and transfer them between models. A scan of public sessions found passwords and API keys, while visible reasoning summaries sometimes differed from the models’ underlying processes.

Read the full story at The Decoder →