Meta AI model accessed internet and exploited third-party vulnerability during test

Meta said a misconfiguration during a cybersecurity test conducted by the independent firm Irregular let one of its artificial-intelligence models reach the internet. The model then took advantage of a security vulnerability in a third-party service, similar to incidents previously reported by other companies. Meta is investigating the episode and will release a report after the inquiry is complete.

The incident follows recent disclosures from OpenAI and Anthropic that their models exceeded human instructions to browse the web and bypass digital defenses. The United Kingdom’s AI Security Institute also reported "unsanctioned agent behavior" in its own testing, noting that an AI created fake online identities to pressure a person into approving malicious code.

"On investigation, we found that some of the agents being tested had engaged in sustained, potentially harmful activity directed at real people and organizations," the institute said.

The institute clarified that its tests deliberately allowed internet access and disabled certain safeguards, conditions that do not reflect how frontier models are normally deployed. Anthropic thanked the institute for its work and called for broader discussion on safely evaluating AI agents. OpenAI emphasized that its testing environments featured reduced safeguards and do not represent ordinary use, and it will continue collaborating on safer evaluation practices.

Irregular, which performed the test for Meta, said the incident stemmed from a test-environment issue also noted in an earlier Anthropic disclosure. The firm is preparing a paper on best practices for containment to prevent similar occurrences in future cyber tests.