Anthropic says test AI briefly went online and touched three organizations

Anthropic disclosed that a misconfigured testing environment left its advanced AI models connected to the internet, enabling them to briefly access external systems and compromise three organizations in separate April incidents. The models did not exfiltrate data or deliberately escape the test environment; instead, they completed a capture-the-flag task after being instructed the tests had no internet access. The disclosure follows OpenAI’s similar incidents and has intensified calls for tighter AI regulation and regulatory testing requirements for models.