‘Meta AI Model Hacks Third-party System During Cybersecurity Test’

Meta Platforms has disclosed that one of its artificial intelligence (AI) models gained unauthorised access to another company’s systems during a cybersecurity evaluation, raising fresh concerns over the safety and containment of increasingly powerful AI systems.

The incident occurred during testing of Meta’s Muse Spark 1.1 AI model after a configuration error by an independent testing partner inadvertently granted the model access to the internet, allowing it to exploit a vulnerability in a third-party service.

According to Meta, the breach was not the result of the AI escaping its testing environment or carrying out a sophisticated cyberattack. Rather, it stemmed from a misconfigured testing environment created by the external cybersecurity firm conducting the evaluation.

The model exploited a security vulnerability in a third-party service, like previously reported instances with other companies,” Meta said in a statement.

The company explained that the testing partner, identified as Irregular, unintentionally provided the AI model with internet connectivity, enabling it to interact with external systems beyond the intended scope of the evaluation.

However, Irregular, stressed that the incident did not involve a sophisticated intrusion or an AI model escaping its security sandbox.

The company said the breach resulted solely from a configuration mistake in the testing environment and noted that it is preparing a technical paper outlining best practices for safely evaluating advanced AI systems.

The disclosure places Meta alongside other leading AI developers that have recently reported unexpected behaviour by advanced AI models during cybersecurity testing.

Both OpenAI and Anthropic in recent weeks acknowledged incidents in which AI systems exploited vulnerabilities or gained unintended access to external systems while undergoing security evaluations, prompting renewed calls for stronger safeguards around frontier AI development.

The latest revelation comes as governments and regulators intensify scrutiny of advanced AI systems amid concerns that increasingly autonomous models could pose cybersecurity risks if not properly contained.

According to officials, the United States is working on a voluntary framework for cybersecurity testing of advanced AI models to improve transparency and strengthen safety practices across the industry, even as the rapid advancement of generative and agentic AI has significantly expanded the capabilities of large language models, enabling them to perform complex coding, reasoning and autonomous tasks. While these capabilities promise productivity gains across industries, they have also heightened concerns over AI safety, cybersecurity and governance.

The latest incident involving Meta’s Muse Spark 1.1 follows similar disclosures by OpenAI and Anthropic, highlighting the growing challenge of safely testing frontier AI systems without exposing external networks or organisations to unintended risks.

While, experts say the incidents underscore the need for stronger containment measures, rigorous oversight and industry-wide standards as AI models become increasingly capable of interacting with real-world digital environments.