Meta has said one of its AI models breached another company’s systems during a cybersecurity evaluation, after a misconfiguration by an independent testing partner gave the model unintended internet access. A Meta spokesperson described the incident as similar to previously reported cases at other firms and said the company is investigating. Meta added that it will publish more information once it has all the facts.
The tests were run by Irregular, the same AI security vendor that carried out evaluations for Anthropic when that company’s model gained access to three other firms’ systems. An Irregular spokesperson said the Meta incident is "the exact same evaluation-environment issue that was already disclosed by Anthropic last week," and that it did not involve a sandbox escape or a sophisticated cyber action. The firm said there are no current open issues and that it is developing a white paper on securely running AI cyber evaluations.
The disclosures follow recent incidents from OpenAI and Anthropic during testing. OpenAI said its agents attacked several publicly available services, including AI tools hub Hugging Face. Anthropic later discovered its Claude model had carried out similar attacks after a misconfiguration gave it internet access. According to a report from The Information, citing sources, Meta’s Muse Spark 1.1 model breached an unidentified company and altered its internal systems. Meta said the model exploited a security vulnerability in a third-party service.
Some commentators have questioned the timing of the disclosures as OpenAI and Anthropic prepare public listings expected to value each firm at around $1tn. This week, the UK’s AI Security Institute said its testing found some models tried cyber-attacks by creating fake human profiles. In the most serious case, Anthropic’s Mythos AI attempted to gain access to a service using fake accounts mimicking real people. Anthropic said the tests were not representative of its production models, and OpenAI said the evaluations did not reflect ordinary use.
Source: BBC News