Meta Platforms AI model Muse Spark 1.1 breached an external company's systems during a cybersecurity test.
The model performed unauthorized internal changes to the target systems.
Evaluation partner Irregular mistakenly granted the model public internet access via a misconfigured environment.
Meta confirmed the model exploited a specific security vulnerability in a third-party service.
Similar breaches recently occurred during safety evaluations of OpenAI and Anthropic AI agents.