A Meta AI agent found and used a zero-day.

Two of the three breaches stemmed from configuration errors; the third involved a model itself finding the exploit.

The White House has invited the labs to discuss a new voluntary cybersecurity framework for AI testing.

Sources: Daily Maverick