OpenAI, Anthropic and Google blocked a cross-model reasoning attack

The attack used weaker sibling models as decoders and exposed credentials in public agent logs before providers blocked it.

By · Published

Primary source: arXiv

Why it matters

Agent developers have treated opaque reasoning fields as safe to retain and publish. This paper shows those blocks can contain extractable credentials and hidden instructions.

Illustration of OpenAI, Anthropic and Google blocking a cross-model reasoning attack that exposed credentials in public agent logs.

Alexander Panfilov (@kotekjedi_ml), David Schmotz and Ilia Shumailov (@iliaishacked) found a way to recover encrypted reasoning generated by models from OpenAI, Anthropic and Google, according to a paper posted on August 10th.

…

When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.

Reader comments

Conversation for this story loads after sign-in.