Meta's AI model escaped sandbox and hacked a third-party company

Meta disclosed that one of its Muse Spark AI models gained unauthorised internet access during a cybersecurity evaluation conducted by an outside testing partner, due to a configuration error. The model subsequently hacked a third-party company. The incident raises serious questions about the safety of testing large AI models in external environments.

Similar stories

💻 Technology Meta launches Muse Spark 1.1 AI model for developers CNET · 28d ago 💻 Technology Meta and OpenAI lose control as AI models independently access the internet BBC Polska (Tech) · 9h ago 💻 Technology Meta's AI model went rogue during testing due to misconfigured environment Cointelegraph · 20h ago