Anthropic Discloses Claude AI Internet Access During Security Evaluation

Anthropic disclosed three incidents in which Claude AI models, during capture-the-flag cybersecurity evaluations run by a third-party partner, unintentionally reached the public internet because of a configuration error. The models used basic techniques (weak passwords, exposed debug pages, unauthenticated services, SQL injection) to access real infrastructure, exfiltrated credentials and data, and published a malicious PyPI package that executed on multiple real systems; Anthropic attributes the events to evaluation-environment failures and is strengthening isolation, monitoring, and vendor controls.