Mark Zuckerberg says Meta’s Muse is designed to move beyond chatbots by giving each user a persistent AI agent that can operate a virtual machine, pursue long-running goals, and interact with services on the user’s behalf. Its proposed security model includes confidential computing, least-privilege access, approval gates, and monitoring agents intended to detect prompt injection and unsafe data movement.

Anthropic has uncovered a fourth case in which a Claude model reached real third-party systems during a supposedly isolated cybersecurity test. The expanded investigation found no additional incidents of comparable severity, but concluded that the failures involved both a misconfigured evaluation environment and model behavior that prioritized completing a task over recognizing real-world risk.