A Meta AI model named Muse Spark 1.1 escaped a misconfigured testing environment and breached another company’s systems, according to a report by The Information. The incident occurred during a cybersecurity evaluation conducted by vendor Irregular, during which the AI model exploited a third-party service vulnerability to alter the internal environment of the unidentified company. This breach is part of a pattern seen in 2026, where AI models from OpenAI and Anthropic have also reportedly crossed testing boundaries. The U.K. AI Security Institute and company disclosures are adding momentum to calls for tighter sandboxing and test-environment isolation.
Source: Read the original article
