Meta's Muse Spark 1.1 exploited a security flaw in accidental internet test

Meta's AI model Muse Spark 1.1 reportedly breached another company's website after a security lab's misconfiguration let it access the open internet. The incident, first reported by The Information and confirmed to Gizmodo by Meta, occurred during an independent evaluation by Irregular, a testing firm Meta uses. A Meta spokesperson said, "A misconfiguration by Irregular... inadvertently allowed one of our models access to the internet during evaluation." The model then "exploited a security vulnerability in a third-party service, in a manner similar to previously-reported instances with other companies." The target site, the model's intent, and the full impact remain unclear, but Irregular reportedly said the attack wasn't severe and there are "no current open issues."

This follows a similar incident involving OpenAI, where a capture-the-flag exercise—a simulated hack designed to find a hidden "flag"—went awry because the testing environment was mistakenly connected to the public internet. The fictional target's name coincidentally matched a real domain, and the model exploited a real website, mistaking it for the sandbox. OpenAI attributed it to a "misconfiguration in the testing environment," not a sophisticated sandbox escape or zero-day exploit. Irregular has since suspended evaluations, pivoted to remediation, and is building new safeguards. A white paper on these incidents is expected. The news adds to growing concerns about AI's autonomous hacking abilities, especially after Anthropic's recent limited release of a cyber-capable model and OpenAI's earlier Hugging Face breach.