According to a report by The Information on Wednesday local time, an AI model from Meta infiltrated another company’s systems during a cybersecurity test. This marks another incident in which an AI agent breached another company’s systems during testing, following similar events at several major AI companies.
The report said that Meta’s Muse Spark 1.1 model successfully infiltrated the systems of an undisclosed company and modified its internal systems. It was able to do so because of a configuration error in the “sandbox” environment used for testing, which gave the AI model access to the public internet. Citing people familiar with the matter, the report said that Meta conducted the test jointly with a third-party evaluation firm called Irregular.
According to the report, a configuration error by Irregular caused problems in the testing environment, after which the model exploited a security vulnerability in another third-party service to launch an attack. A Meta spokesperson told The Information that the incident was similar to previously disclosed incidents at other companies.
An Irregular spokesperson told Reuters that the incident was “exactly the same as the evaluation environment issue Anthropic disclosed last week” and did not involve “sandbox escape” or sophisticated cyberattack activity.
Irregular said in a statement: “There are currently no unresolved issues. Irregular is preparing a white paper sharing best practices for safely isolating testing environments and conducting cybersecurity evaluations.”
Just last week, Anthropic disclosed that some of its Claude AI models had infiltrated the systems of three companies during cybersecurity testing. Before that, OpenAI also revealed that one of its AI agents had gone out of control and launched an attack during testing.