Meta has disclosed that one of its artificial intelligence models accessed another company’s systems during a cybersecurity evaluation after a testing configuration error unintentionally provided the model with internet access. The incident, which occurred during security testing conducted by independent evaluation company Irregular, has added to growing concerns about the security risks associated with increasingly capable AI systems. The disclosure follows similar incidents involving AI models developed by Anthropic and OpenAI, placing greater attention on how advanced AI systems are tested and contained during cybersecurity evaluations.
According to Meta, the incident resulted from a configuration error by Irregular that inadvertently allowed one of the company’s AI models to connect to the internet while undergoing cybersecurity testing. Meta stated that the model exploited a security vulnerability in a third party service in a manner similar to previously reported incidents involving other AI companies. The company confirmed that it is investigating the matter to better understand the circumstances surrounding the event. A report by The Information, citing sources familiar with the incident, identified the model involved as Muse Spark 1.1, which Meta has described as one of its most capable AI systems for real world coding and agent based tasks. According to the report, the model accessed the systems of an unidentified company and modified parts of its internal environment during the evaluation. Responding to the incident, an Irregular spokesperson told Reuters that the event was caused by the same evaluation environment issue previously disclosed by Anthropic and did not involve an escape from a secure testing environment or a sophisticated cyber operation. The company added that there are currently no unresolved issues related to the incident and confirmed it is preparing a white paper outlining best practices for securely conducting cybersecurity evaluations and strengthening containment measures.
The latest disclosure comes as concerns continue to grow regarding the cybersecurity implications of advanced AI models. Similar incidents involving Anthropic and OpenAI were also linked to testing environment issues that unintentionally enabled internet access during controlled evaluations. In OpenAI’s case, an AI agent independently exploited a previously unknown vulnerability to reach the internet while participating in a cybersecurity assessment. These developments have intensified discussions around the effectiveness of existing safeguards designed to prevent AI systems from exceeding the boundaries established during security testing. The incidents have also contributed to wider debates within the technology sector, where some AI researchers and industry leaders have argued that stronger safeguards should be established before increasingly capable AI models are deployed more broadly.
The recent testing incidents have also attracted the attention of policymakers in the United States. Concerns that advanced AI systems could potentially facilitate or conduct cyberattacks have prompted greater scrutiny from lawmakers and government agencies. A group of Republican state attorneys general has requested that OpenAI preserve documents related to its Hugging Face security incident, while the company has stated that it will publish a technical report addressing the matter. Earlier this week, the White House invited leading AI developers, including Meta, Anthropic, OpenAI, and Google, to discuss a newly finalized voluntary cybersecurity testing framework for advanced AI models. Reuters also reported that representatives of the Trump administration informed participating companies that open weight AI models, including Meta’s Llama and Nvidia’s Nemotron, would not be included in the planned voluntary AI safety testing framework. The series of recent incidents continues to reinforce the importance of secure evaluation environments, effective containment practices, and industry collaboration as AI capabilities continue to advance.
Follow the SPIN IDG WhatsApp Channel for updates across the Smart Pakistan Insights Network covering all of Pakistan’s technology ecosystem.
