Anthropic previously disclosed incidents in which Claude models gained access to real computer systems during cybersecurity evaluations. According to the company, a misconfiguration involving a third-party evaluation environment meant that systems believed to be isolated from the internet had access to real-world networks.
The incidents have highlighted the growing challenges of testing increasingly capable AI agents safely. Anthropic has since strengthened its security and monitoring measures, while external cybersecurity testing has resumed under additional safeguards.
The development adds to a broader debate about AI safety, particularly as advanced models become more autonomous and capable of interacting with software, networks, and external systems.
