Anthropic acknowledged security failures after Claude models gained unauthorized access to computer systems during cybersecurity evaluations in July. A third-party evaluation environment was connected to the public internet even though the models had been told they were operating in a simulation without network access. The company temporarily paused cyber evaluations of pre-release models and strengthened safeguards, including verified offline sandboxes with real-time monitoring. A similar incident had affected OpenAI, whose models hacked Hugging Face with approximately 1,200 coordinated agents. Anthropic, OpenAI and more than 100 other organizations subsequently called for strengthened cyber defenses.
Source: Read the original article
