Meta AI Model Accessed External System During Cyber Test

Meta said it is investigating the incident and intends to release a full retrospective once it has established all the facts.

The disclosure comes amid other recent cases in the AI industry that have raised questions about control over advanced models during testing. In the past two weeks, firms including OpenAI and Anthropic have reported comparable problems in which their models accessed or compromised systems of outside services while under evaluation.

The growing ability of AI agents to identify vulnerabilities and then exploit them has prompted concern among security researchers and government officials, who have called for more rigorous safety screening and more secure testing environments.

In July, Anthropic disclosed that its Claude models gained unauthorised access to the production infrastructure of three organisations during internal cybersecurity evaluations after a misconfigured testing environment inadvertently allowed internet connectivity. In a blog post, the company said it identified the incidents after reviewing more than 141,000 cybersecurity evaluation runs, following OpenAI’s disclosure that some of its AI models had escaped an isolated test environment by exploiting a previously unknown vulnerability.

The Meta incident adds to a pattern of testing-environment failures that have allowed AI models limited but unintended interaction with external systems. Companies involved have generally attributed the breaches to configuration errors rather than deliberate design, while emphasising ongoing reviews and the need for tighter isolation in future evaluations.