AI agent failure taxonomies
6 published taxonomies of AI agent failures and risks, kept as first-class data, with explicit crosswalks from each category into SAICA's coding-agent failure modes. An archived cluster study from April 2026 compared an earlier version of them; it is not maintained.
-
9 Critical Failure Patterns of Coding Agents
v2026-01 · 9 categoriesColumbia University DAPLab (blog post by Reya Vir)
Nine failure patterns that the authors saw while building 15+ applications with five coding agents and vibe-coding tools (Claude, Cline, Cursor, v0, Replit; testing as of November 2025). The post says it documented "hundreds of failures" during development and grouped them into the nine patterns. It is a blog post about building web applications, not a peer-reviewed study, and it gives no counts per pattern.
-
Characterizing Faults in Agentic AI: A Taxonomy of Types, Symptoms, and Root Causes
varXiv v2 (2026-05-07) · 76 categoriesShah, Morovati, Rahman, Khomh (Dalhousie University; Polytechnique Montréal)
Taxonomy of faults in agentic AI software, built from 13,602 closed issues and merged pull requests in 40 open-source repositories; 385 faults were sampled and coded by hand, and 145 practitioners rated the result. Version 2 of the paper has three lists: 34 fault types in 14 categories and four dimensions (Figure 2), 12 symptom categories (Table 1) and 12 root-cause categories (Table 2). The unit of study is a defect in the code of an agent framework, library, tool or application. It is not a failure of a coding agent on a user's task.
-
MAST: Multi-Agent System Failure Taxonomy
varXiv 2503.13657v3 (2025-10-26) · 17 categoriesCemri et al. (UC Berkeley + collaborators)
Failure modes of multi-agent LLM systems: 14 failure modes in 3 categories, derived from expert annotation of 150 traces (kappa = 0.88) and applied to MAST-Data (1,600+ annotated traces across 7 multi-agent frameworks).
-
OWASP Top 10 for Agentic Applications 2026
v2026 · 10 categoriesOWASP Gen AI Security Project
Ten security risks of autonomous and agentic AI applications, ASI01 to ASI10, published by the OWASP Gen AI Security Project (the document is marked "Version 2026" and dated December 2025). A security list: each entry is a class of attack or weakness, with examples, attack scenarios and mitigations. It is not limited to coding agents.
-
Swiss Cheese Model for AI Safety: Multi-Layered Runtime Guardrail Taxonomy
v2025-01 (arXiv v4; IEEE ICSA 2025) · 44 categoriesShamsujjoha, Lu, Zhao, Zhu (CSIRO Data61)
A taxonomy of runtime guardrails for foundation-model-based agents, built from a systematic literature review. The paper names three dimensions: quality attributes, pipelines and artifacts. It also lists design options for guardrails; of those, this file holds the guardrail actions. The paper uses the taxonomy to propose a multi-layered "Swiss Cheese" reference architecture. The categories describe guardrails, not failures.
-
Taxonomy of Failure Mode in Agentic AI Systems (Microsoft AI Red Team, version 1.0)
v1.0 (2025-04) · 27 categoriesMicrosoft AI Red Team (AIRT)
Version 1.0 of the Microsoft AI Red Team's list of failure modes of agentic AI systems: 27 modes on two axes, safety or security and novel (only in agentic systems) or existing (known from other AI systems, with higher risk in agents). 4 novel safety, 6 novel security, 7 existing safety and 10 existing security modes. The paper calls its method a failure mode and effects analysis of current and envisaged systems. It is about agentic systems in general, with many multi-agent cases, not about coding agents.