Taxonomy of Failure Modes in Agentic AI Systems
- Publisher: Microsoft
- Status:
final - Version:
1 - Release Date:
2025-04-24 - Date Added:
2025-12-04 - Source URL: https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/final/en-us/microsoft-brand/documents/Taxonomy-of-Failure-Mode-in-Agentic-AI-Systems-Whitepaper.pdf
Summary
The whitepaper from Microsoft (via its AI Red Team) presents a structured taxonomy of how “agentic” AI systems can fail, categorizing both safety and security risks. The guidance is aimed at engineers, security and safety professionals, and enterprise-governance stakeholders involved in building, evaluating, or deploying agentic AI systems; its purpose is to help them proactively identify, reason about, and mitigate potential failure modes before deployment.
Key Takeaways
- Taxonomy divides failures along two dimensions: whether they affect safety or security, and whether they are novel to agentic AI or are existing AI failures amplified by agentic contexts.
- Traditional generative AI failure modes become more serious when embedded in an agentic system, because the agent can act autonomously, persist state in memory, and make decisions or effect changes in the environment
- Useful across roles: engineering, security, governance
- Intended as a starting point, as authors acknowledge the field is evolving rapidly and new classes of failure may emerge
Additional Sources
- New whitepaper outlines the taxonomy of failure modes in AI agents — blog post introducing and describing the guidance
- Microsoft's Taxonomy of Failure Modes in Agentic AI Systems - Top 10 Insights - breakdown of top failure modes and mitigations derived from taxonomy
License
Proprietary