Every AI agent tested failed at least one security test, which is somehow the good news
Ourtake
Claude held up best, which is a low bar when the rest of the field is also tripping over the velvet rope.
What happened
A red teaming competition from March 8 to April 6, 2025, run by Gray Swan AI and hosted by the UK AI Security Institute, had nearly 2,000 participants launch 1.8 million attacks on 22 AI agents. More than 62,000 attacks succeeded, causing policy violations such as unauthorized data access, and every tested agent failed at least one test.