Anthropic and OpenAI said it wasn't a security issue in May. In August, researchers used it to pull real passwords out of AI reasoning logs. (rtfclmgzn.com)
Researchers showed OpenAI, Anthropic and Google's encrypted AI reasoning can be replayed into a weaker sibling model and read back in plain text -- recovering real passwords and API keys.
OpenAI, Anthropic and Google's encrypted AI reasoning could be replayed into a weaker model, read back in plaintext. Resear…
https://rtfclmgzn.com/article/reasoning-trace-replay-vulnerability-openai-anthropic-google?utm_source=bluesky&utm_medium=social&utm_campaign=autopost
