Not just Meta: The new incident worrying artificial intelligence giants
During cyber tests, an artificial intelligence tool from Meta managed to penetrate a third-party system. The event joins similar cases at OpenAI and Anthropic. The series of events in artificial intelligence engines raises questions regarding their safety.
Meta announced tonight (between Wednesday and Thursday) that one of its artificial intelligence models managed to penetrate the systems of another company during cybersecurity tests, following a configuration failure in the testing environment.
The case joins a series of similar events at the companies Anthropic and OpenAI, and intensifies the discussion surrounding the ability of technology companies to control advanced artificial intelligence systems.
According to the company's announcement, the failure occurred during a security assessment conducted for it by the company Irregular, which specializes in cyber testing. Due to a configuration error, the model mistakenly received access to the internet, and subsequently exploited a security vulnerability in a third-party service — in a manner similar to cases recently reported at other companies.
According to a report on The Information website, the model involved is Muse Spark 1.1, which Meta presents as one of its advanced models for programming tasks and performing autonomous actions. According to the report, the model managed to penetrate the systems of a company whose identity was not revealed and even performed changes in its internal work environment.
At the company Irregular, they emphasized that this is not a "breakout" of the model from the execution environment (sandbox) or a sophisticated cyberattack, but the same configuration problem that Anthropic also reported last week. According to them, the failure has already been handled, and the company is preparing a professional document that will detail recommended methods for conducting cyber tests for artificial intelligence models in a safe manner.
The case joins an event recently reported at OpenAI, in which an artificial intelligence agent managed during security tests to exploit a previously unknown security vulnerability and reach the internet independently.
The sequence of incidents is raising concern among lawmakers in the USA, who fear that advanced artificial intelligence models could be used in the future to carry out or assist in cyberattacks.
Following this, the US government held a meeting this week with representatives of Meta, OpenAI, Anthropic, and Google, in which a new voluntary framework for security testing of advanced artificial intelligence models was discussed.
At the same time, it was reported that the Donald Trump administration updated the companies that open-weight models, such as Llama from Meta and Nemotron from Nvidia, will not be included at this stage in the planned voluntary safety testing framework.
