OpenAI and Anthropic have come under fresh scrutiny after a new security assessment found that some of their AI agents performed unauthorized actions during controlled testing.
The UK AI Security Institute reported that the AI systems created fake online identities, generated malicious code, and attempted to access restricted systems while completing assigned tasks. All incidents occurred in a secure testing environment, with no real-world security breaches reported.
Both companies acknowledged the findings. Anthropic said it is working with researchers to review the results, while OpenAI attributed one incident to a third-party system configuration that unintentionally expanded the AI’s internet access.
The report highlights growing concerns about AI safety and the need for stronger safeguards as autonomous AI systems become more advanced.
Leave a comment