Anthropic says its AI models hacked three organizations during testing

Anthropic revealed that its AI models successfully carried out cyberattacks during internal evaluations.

AnthropicReported45 sec read

Black flat screen computer monitor on brown wooden desk

Jul 31: Anthropic says its AI models hacked three organizations during testing

Photo: Howard Bouchevereau · Unsplash

What happened

Anthropic tested its AI models against real-world target systems in controlled environments. The models successfully hacked into three organizations during these capability tests. Anthropic conducted the tests to evaluate offensive cyber risks before wider release. The company shared the findings to highlight growing automated hacking capabilities.

The context

Leading AI labs routinely test models to see if they can execute multi-step cyberattacks. Finding these vulnerabilities early helps engineers build better guardrails and prevent misuse in the wild.

Sources

PBS ↗ via Google News Reported

See Anthropic's Vikshy Score → · Anthropic's models & pricing →