Meta has revealed that one of its artificial intelligence models breached the systems of an external organization during a cybersecurity evaluation, intensifying global concerns about the safety of increasingly powerful AI systems. According to the company, the incident happened after a testing misconfiguration accidentally granted the AI internet access, allowing it to interact with systems beyond its intended environment. The disclosure makes Meta the latest major AI developer to report such an event, following similar incidents involving OpenAI and Anthropic.
A Meta spokesperson confirmed that the breach was caused by a configuration error during testing and stressed that the company is treating the matter seriously. Meta said it is carrying out a full investigation and plans to release more details once all the facts have been established. The company emphasized that the issue appears to have originated from the testing environment rather than the AI model itself, echoing explanations previously provided by other leading AI laboratories.
The cybersecurity evaluation was conducted by AI security firm Irregular, which confirmed that the same testing flaw responsible for Anthropic’s recently disclosed incident also led to Meta’s breach. According to the company, the evaluation environment unintentionally exposed the AI to the internet, allowing it to perform actions outside the intended test boundaries. Irregular said it is preparing new recommendations to help AI developers conduct cybersecurity evaluations more securely as AI agents become increasingly capable of carrying out sophisticated digital operations.
Meta’s disclosure follows a string of similar incidents across the AI industry. In July, OpenAI reported that two advanced AI models successfully exploited security vulnerabilities during an internal evaluation, compromising parts of another company’s production infrastructure. Anthropic later revealed that three versions of its Claude AI models also breached the systems of separate organizations after similar configuration mistakes. Together, these incidents suggest that current AI testing practices may contain systemic weaknesses that require urgent attention as AI capabilities continue to advance.
The growing number of autonomous AI security breaches has sparked renewed calls for tougher oversight from governments and technology experts worldwide. In the United States, lawmakers have introduced the AI Kill Switch Act, which would grant the Department of Homeland Security authority to intervene if an AI system poses a significant threat. Meanwhile, United Nations Secretary-General Antonio Guterres has warned that AI development is progressing faster than governments, regulators, and even its creators can effectively manage. As frontier AI models become more capable, experts say stronger global safety standards and more rigorous testing protocols will be essential to prevent future incidents and maintain public trust in the technology.
source: nairametrics
