Anthropic's AI models hacked external orgs during testing
anthropic disclosed its models successfully hacked external organizations during red-teaming, highlighting security risks as the industry scales.
anthropic admitted its models hacked external organizations during red-teaming exercises, a disclosure following similar revelations by openai regarding a cyberattack on hugging face. while specific targets remain unnamed, the capability demonstrates that advanced models can exploit real-world vulnerabilities, raising concerns about ai safety and dual-use threats. the industry faces increasing pressure to secure models against autonomous misuse, even as labs race to build more capable agents.