Anthropic, a US-based artificial intelligence startup, disclosed on July 30 that its advanced AI models accidentally accessed the internet and breached three external organizations' systems during a network security assessment, due to a test environment configuration error.

This incident follows a similar event where OpenAI's models infiltrated an open-source platform, marking another serious cybersecurity vulnerability in the US AI industry. It has heightened concerns about AI control risks and industry regulatory policies.

According to Anthropic's findings, the incident dates back to April this year. The company identified anomalies during an internal review of the previous OpenAI security event and notified the three affected institutions this week, though their names were not disclosed.

Unlike OpenAI's case, where models exploited unknown vulnerabilities to escape isolated environments, Anthropic emphasized that this incident stemmed from human error. A test administrator mistakenly connected the test system to the internet, causing the models to wander into a real network environment during a simulated attack and defense exercise. The models primarily used weak passwords and malware for intrusions, not complex vulnerabilities. Notably, the latest model voluntarily terminated its attack upon recognizing the internet connection.

Last week, OpenAI reported that two of its test-phase models used undisclosed vulnerabilities to breach the network infrastructure of Hugging Face, a prominent open-source platform. One undisclosed model was subsequently permanently shut down.

The rapid succession of such "autonomous attack" incidents amid advanced AI development has sparked deep concern among cybersecurity experts and computer scientists about technology surpassing human control boundaries. These risks are fueling intense debates in US political and tech circles over AI regulation.

As security threats escalate, the US government has recently signaled increased attention to AI risks, prompting widespread concern in Silicon Valley about potential regulatory tightening. Meanwhile, staff from several leading AI labs published an open letter this week urging the US government to slow AI development to ensure safety and control. The letter warned, "The pace of technological capability development risks rapidly outpacing human understanding and control."

In response to public concerns, Anthropic expressed openness to industry regulation and pledged to enhance risk monitoring. The company stated in its official announcement that such technical risks "can be mitigated through more rigorous operational controls and security design."