Microsoft Identifies 7 New Agentic AI Failure Modes as MCP and Computer-Use Risks Grow
Updated
Updated · InfoWorld · Jun 5
Microsoft Identifies 7 New Agentic AI Failure Modes as MCP and Computer-Use Risks Grow
3 articles · Updated · InfoWorld · Jun 5
Summary
Seven new failure modes were added to Microsoft’s agentic AI taxonomy, expanding its list of ways autonomous systems can be manipulated or leak sensitive internal details.
Microsoft said the update was driven by four shifts: rapid mainstream adoption, a more mature Model Context Protocol ecosystem, the rise of computer-use agents, and more real-world evidence from researchers.
The new risks include supply-chain compromise through natural language, goal hijacking, inter-agent trust escalation, visual attacks on GUI-based agents, session context contamination, MCP or plugin abuse, and architecture disclosure.
Microsoft urged security teams to generate an SBOM for every deployed agent, verify agent identity with attestable credentials, add the seven modes to red-team testing, and audit human-in-the-loop workflows as a security control.
As AI agents become 'digital insiders,' how do we prevent them from becoming the ultimate insider threat?
The Miasma worm just attacked AI assistants. Are security tools prepared for natural language threats?
Seven New Agentic AI Failure Modes Revealed: Microsoft’s 2026 Security Update and the Escalating Risks for Enterprises
Overview
In June 2026, Microsoft released a major update to its agentic AI failure modes framework, reflecting the rapidly evolving threat landscape. This update, built on 12 months of intensive red teaming by the Microsoft AI Red Team, introduces seven new failure modes and expands the taxonomy of AI vulnerabilities. The findings are practical and actionable, serving as a vital reference for AI security practitioners. Notably, the update documents critical zero-click Human-in-the-Loop (HitL) bypass chains, highlighting new ways attackers can compromise agentic systems without user interaction. These insights are essential for anyone building or defending modern AI systems.