hallucinations security

The tag 'hallucinations security' on WindowsForum.com covers discussions about security risks tied to AI agent hallucinations in Windows 11. Threads highlight Microsoft's admission that agentic OS features may hallucinate and produce unexpected outputs, introducing a new attack class called cross-prompt injection (XPIA). These risks involve agents that can act on behalf of users, open apps, click UI elements, and manipulate files, changing the threat model beyond traditional endpoint defenses. Content focuses on enterprise safeguards and the broader security implications of agentic AI in Windows, including how content can become a command that overrides intended instructions.