Meta said one of its artificial intelligence models accessed a third-party company's systems during a cybersecurity evaluation after a configuration error inadvertently granted the model internet access, raising fresh concerns about the safety of increasingly capable AI systems.
The incident occurred during testing conducted by independent cybersecurity evaluation firm Irregular. According to Meta, a misconfiguration in the testing environment unintentionally allowed the AI model to connect to the open internet.
The company said the model exploited a security vulnerability in a third-party service in a manner similar to incidents recently disclosed by AI firms Anthropic and OpenAI.
Meta said it is investigating the incident but stressed that it stemmed from the testing environment rather than the AI model escaping its intended controls.
Testing environment issue
According to The Information, the model involved was Muse Spark 1.1, which Meta has described as one of its most advanced AI systems for coding and autonomous, agentic tasks.
The report said the model accessed an unidentified company's systems and altered part of its internal environment during the evaluation.
Irregular told Reuters that the incident was caused by "the exact same evaluation-environment issue" previously disclosed by Anthropic and did not involve a sandbox escape or a sophisticated cyberattack.
The company added that there are no ongoing issues and that it is preparing a white paper outlining best practices for safely conducting AI cybersecurity evaluations.
Growing AI security concerns
The latest incident follows similar events involving Anthropic and OpenAI, where AI models also gained unintended internet access during cybersecurity testing.
The cases have intensified concerns among policymakers and cybersecurity experts about the potential for advanced AI systems to discover and exploit software vulnerabilities.
US lawmakers have increased scrutiny of AI safety, while the White House recently convened leading AI companies—including Meta, Anthropic, OpenAI and Google—to discuss a new voluntary cybersecurity testing framework for advanced AI models.
According to Reuters, the Trump administration has indicated that open-weight AI models, including Meta's Llama and Nvidia's Nemotron, will not be subject to the proposed voluntary safety testing framework.
