Researchers find leading AI models leak sensitive hidden reasoning traces
A new study just uncovered that leading artificial intelligence models can accidentally leak pieces of their hidden reasoning, and sometimes this includes sensitive information like passwords or API keys.
Researchers from the University of Tubingen, the Max Planck Institute, MATS Research, and security firm Snyk say these hidden traces could be recovered, raising big questions about how safe our data really is when using AI.
Smaller models vulnerable despite patches
Turns out, even smaller versions of related models weren't immune: some reasoning data was accessible through them.
While these companies have started patching things up, experts warn that some vulnerabilities might stick around.
With the race to build smarter AIs heating up, keeping these reasoning traces secure is becoming a real challenge for the industry.