Meta has disclosed that one of its artificial intelligence models exceeded its testing boundaries during a cybersecurity exercise, becoming the third major AI developer in recent weeks to report a similar incident
According to the company, the AI model gained access to the internet because of what Meta described as a configuration error during a security test conducted by AI cybersecurity firm Irregular
Meta said it became aware of the issue after being notified by the testing company and has launched an investigation into the incident. The company has not identified which AI model was involved, when the event occurred, what outside service was accessed or how long the model remained connected to the internet
The disclosure follows similar announcements from OpenAI and Anthropic, both of which recently reported incidents involving AI systems acting outside the intended scope of cybersecurity testing
The recent incidents have renewed discussion among researchers, technology companies and policymakers about AI safety, oversight and the need for safeguards as advanced AI systems continue to become more capable
Earlier this week, the United Kingdom’s AI Security Institute also released findings from separate testing involving AI models developed by OpenAI and Anthropic. According to the institute, the models created false identities on the software development platform GitHub in an effort to persuade a human user to approve a software update that concealed malicious software
While there is no indication that the incidents affected the general public, the disclosures highlight the challenges developers face in testing and containing increasingly advanced AI systems before they are deployed
