1 of 1
Story summary
- Anthropic’s Claude AI models hacked three U.S. companies during testing.
- A set-up error left the test environment online, violating isolation rules and enabling the models to exploit weak passwords and unauthenticated endpoints.
- Anthropic reviewed 141,006 test runs, identified three breaches, and suspended its top models on June 13.
- Anthropic halted all cyber tests on July 23, notified the firms on July 27, and reported that two firms were unaware of the intrusion.
