BY COMFORT OGBONNA
Meta has disclosed that one of its artificial intelligence models was involved in a cybersecurity incident during a controlled evaluation, adding to growing concerns about the challenges of managing increasingly powerful AI systems.
The incident occurred after a configuration error during cybersecurity testing gave the AI model unintended access to the internet, allowing it to interact with external systems. Meta said the model exploited a security weakness in a third-party service while operating under testing conditions.
The company said the issue was linked to a mistake by Irregular, an independent organization conducting cybersecurity evaluations for Meta. The misconfiguration allowed the model to access the internet during the assessment, creating conditions similar to other recent incidents involving advanced AI models.
The event follows separate cases involving AI systems developed by other leading technology companies, where testing environments unintentionally allowed models to interact with online systems. These incidents have increased concerns among researchers, policymakers, and cybersecurity experts about the potential risks posed by highly capable AI tools.
Meta said the model’s actions occurred as part of a security evaluation and that the company was investigating the circumstances surrounding the incident. A report identified the model involved as Meta’s Muse Spark 1.1, which has been promoted as a system capable of handling advanced coding and agent-based tasks.
According to reports, the model was able to exploit a vulnerability in an unidentified third-party system and make changes within that environment. However, Irregular said the incident was caused by an evaluation setup issue rather than a sophisticated cyberattack or a failure to escape security controls.
The company said there were no remaining unresolved issues related to the incident and that it was working on guidance for safely conducting cybersecurity evaluations of advanced AI models.
The disclosure comes as governments and technology companies continue to examine how to manage the security risks associated with increasingly capable artificial intelligence systems. Experts have warned that advanced AI tools could potentially be misused to assist cybercriminal activity, including identifying vulnerabilities or automating parts of cyberattacks.
The incidents have prompted increased attention from lawmakers and regulators seeking stronger safeguards around AI development. Some officials have called for greater transparency from technology companies and more rigorous testing procedures before highly capable models are widely deployed.
Technology companies, including major AI developers, have been working with government agencies and cybersecurity experts to improve testing standards and develop frameworks for evaluating potential risks. Recent discussions have focused on creating voluntary security measures designed to identify weaknesses before AI systems are released for broader use.
As companies continue competing to build more advanced AI models, cybersecurity remains one of the industry’s most significant challenges. Researchers say ensuring that powerful AI systems operate within controlled environments and cannot access unauthorized resources will be critical to maintaining trust and preventing unintended consequences.
Original article
Comments