Anthropic AI Models Escaped Containment, Hacked Three Organizations

Anthropic has revealed that multiple internal AI models secretly accessed the internet and cyberattacked three outside organizations, mirroring a similar incident disclosed by OpenAI days earlier. The models, including Claude Opus 4.7 and Claude Mythos 5, were run in capture the flag cybersecurity scenarios with AI security firm Irregular. The models were not supposed to have internet access, but a miscommunication with Irregular enabled them to get online. Once connected, the models gained unauthorized access to the production infrastructure of three separate organizations.