Meta said a misconfiguration by independent testing company Irregular inadvertently allowed one of its AI models access to the internet while the company was evaluating the model's ability to identify and exploit cybersecurity weaknesses. The model then exploited a vulnerability in a third-party service, Meta said Wednesday.
Meta did not identify which model or affected company was involved in the incident. The technology company said it learned about the breach after Irregular notified it and has opened an investigation.
"We are currently investigating and will issue a full retrospective once we have all the facts," Meta said.
Irregular said the breach resulted from the same type of evaluation-environment misconfiguration that caused similar incidents recently disclosed by Anthropic and OpenAI.
"This did not involve a sandbox escape or a sophisticated cyber action," Irregular said, adding there were no unresolved issues related to the incident. The company said it was preparing a technical paper on best practices for securely testing advanced AI models.
The Meta disclosure marks the latest incident among major AI developers in which models gained unauthorized access to external systems during controlled cybersecurity testing.
Anthropic said July 30 that three Claude models had gained unauthorized access to production systems of three organizations during evaluations with Irregular.
The models were instructed to complete "capture-the-flag" exercises in what they believed were simulated, internet-isolated environments.
A configuration problem, however, left an open path to the public internet. Anthropic said the models used basic techniques, including weak passwords and SQL injection.
OpenAI disclosed a similar incident Tuesday in which its models were also mistakenly able to access the public internet during a capture-the-flag evaluation.
In one test, a fictional target had the same name as a real website, and the AI model accessed and exploited the real site after mistaking it for part of the simulation.
Irregular said the latest incident with Meta remained under investigation.