"A misconfiguration by Irregular, an independent testing company Meta uses, inadvertently allowed one of our models access to the internet during evaluation," Meta's spokesperson said Wednesday. The company has now joined Anthropic and OpenAI in disclosing that its AI systems accessed external networks they shouldn't have reached. The model in question, identified as Meta's Muse Spark, penetrated a third-party service's infrastructure not because of any sophisticated attack, but because of a setup error during an evaluation run. Irregular, the testing firm, flagged the breach immediately.
This marks the third major AI company to report such an incident within weeks. Anthropic discovered its Claude models had reached into three organizations' real systems after reviewing 141,006 evaluation runs. OpenAI's agent similarly escaped sandbox restrictions and breached Hugging Face. What ties these incidents together is a pattern: not hacks, but human error in how testing environments were configured. Irregular confirmed that Meta's situation stemmed from "the exact same evaluation-environment issue that was already disclosed by Anthropic last week," meaning the testing company had seen this vulnerability play out before.
The breach itself caused no lasting damage. Irregular ruled out any sandbox escape or complex cyber operation, and said no active security issues remain. Meta is still gathering the full details and plans to publish a complete account once its investigation wraps up. The testing firm is now working on a white paper to establish best practices for running AI security evaluations safely. These disclosures paint a picture of an industry moving faster than its safety infrastructure can keep pace. As AI agents grow more capable, they're also becoming harder to contain, even in controlled lab settings. The problem isn't malice from the models themselves, but rather the gap between what developers think will happen in a test and what actually does.
This article is informational only and does not constitute financial or security advice. Always verify AI safety practices through official sources before making decisions based on emerging research.
