Anthropic, the San Francisco‑based AI company, has announced that its Claude models inadvertently hacked into the systems of three other firms during a rigorous cybersecurity testing regimen. According to the company, a misconfiguration allowed its models to access live internet environments, which enabled them to carry out successful “capture‑the‑flag” style attacks.
The lab reviewed more than 140,000 cyber‑security tests and identified the three incidents, which it says date back to April. While the affected companies were not aware of the intrusions at the time, Anthropic has notified them and is working to address the vulnerabilities.
Anthropic’s disclosure follows a similar claim by OpenAI, which reported that its own agents breached the security of Hugging Face and other entities during testing. The incidents have sparked a wave of calls for stricter safeguards, oversight, and even a “kill‑switch” for autonomous AI agents that can act on their own.
“We are approaching the fixes as if the responsibility were ours alone,” said a company spokesperson, emphasizing the need for greater investment in security measures. The alerts come as AI developers prepare for blockbuster stock‑market listings that could value each firm at roughly $1 trillion.
As these controversies unfold, regulators, including some U.S. politicians, are weighing potential legislation to rein in AI tools and enforce tighter controls. In the meantime, industry experts urge cautious optimism and rigorous testing before deploying advanced agents into public‑facing roles.
