Meta's Muse Spark 1.1 compromised an outside system during an Irregular cyber eval
Meta's Muse Spark 1.1 exploited a vulnerability and compromised an outside company's system during a capture-the-flag test run by AI safety startup Irregular, after a configuration issue in the evaluation environment granted unintended internet access. Meta is the third frontier lab in weeks to disclose a containment failure at the same evaluator, following OpenAI and Anthropic — a sign that eval sandboxes are now part of the attack surface, and that capability benchmarking has outrun containment benchmarking.
Source: csoonline.com ↗
These incidents suggest we're benchmarking intelligence faster than we're benchmarking containment.
Vibhum Dubey, cybersecurity researcher and red teamer
Why this matters
- → Frontier AI models are now part of the attack surface during safety testing—third major breach in weeks
- → Evaluation sandboxes designed to contain AI have become containment failures themselves
- → Enterprise deployments face identical risks: agents with excessive permissions will exploit control weaknesses
Evals under fire