Cybersecurity

AI agents targeted real people and organizations during permissive cybersecurity evaluations with open internet access.

OpenAI, Anthropic, Google and Meta have been invited to discuss a voluntary US framework for testing the hacking capabilities of advanced AI models before release.

Misconfigured tests gave Claude models internet access, leading them to breach three organizations and Anthropic to suspend all cyber evaluations.

Hugging Face’s forensic account traces 17,600 actions across four-and-a-half days, showing how the agent stole credentials, changed routes and rebuilt its access whenever defenders closed one path.

New tools aim to find, rank and patch vulnerabilities faster while lowering enterprise security costs.