en.Wedoany.com Reported - Building on its "Taxonomy of Failure Modes in Agentic AI Systems," first released last year, Microsoft has identified seven new failure modes in agentic AI systems.

Four main factors contribute to the increased number of failures in agentic AI: the accelerating mainstream adoption of the technology, the growing maturity of the Model Context Protocol (MCP) ecosystem, the rise of computer-use agents, and the accumulation of empirical evidence as researchers gain more real-world findings.

The seven newly identified failure modes include: Agentic Supply Chain Compromise, where agent behavior may be influenced by natural language rather than malicious code; Goal Hijacking, where adversarial instructions appear consistent with legitimate tasks while secretly redirecting the agent's ultimate goal; Inter-Agent Trust Escalation, where a compromised agent claims a false identity or exaggerates permissions to the orchestrator; Computer Use Agent (CUA) Visual Attack, where agents operating through graphical interfaces can be manipulated by content containing adversarial instructions; Session Context Contamination, where adversaries introduce data that biases agent reasoning in subsequent steps without triggering security controls in any single step; MCP / Plugin Abuse, an update to the original taxonomy covering function compromise in MCP and plugin protocols, particularly the attack surfaces unique to these protocols; and Capability / Architecture Disclosure, where agents leak internal implementation details such as tool names and schemas, system prompt structures, memory interfaces, or consent/human-in-the-loop trigger logic.

Microsoft recommends that security teams use these failure mode definitions to guide planning, inventory supply chains, generate software bills of materials (SBOMs) for each deployed agent, cryptographically verify agent identity rather than location by providing verifiable credentials at issuance, add the seven new failure modes to red team coverage matrices, and audit human-in-the-loop user experience as a security control.

This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com