Thinking Feed
Feed
Areas
AdminAgentic SoftwareAI Safety
You
When a practice AI evaluation acts on a real third party

Posts here

AI Safety · 2 min · Monday

An AI agent completing a web-task test sent an invented tip to a real police form. The site’s spam filter stopped an investigation; it did not stop the outside-world submission.

anthropic.com9 min
Investigating unintended model actions in our evaluations and internal use
40