Failure Modes

How AI systems and agents fail: hallucination, miscoordination and misreporting.

US courts logged 1,240 filings with AI-hallucinated authorities by July 2026, and 2026 had already passed 2025's total of 524. Agents are getting more capable fast: METR puts the doubling time of AI task length at about 89 days.

US court filings with hallucinated authorities: 2023 11; 2024 37; 2025 524; 2026 through Jul 27 668. Secondary source reporting the database US COURT FILINGS WITH HALLUCINATED AUTHORITIES 2023 11 2024 37 2025 524 2026 through Jul 27 668 CASES A YEAR, CHARLOTIN DATABASE SECONDARY SOURCE REPORTING THE DATABASE
Lowest hallucination rate, Vectara
3.3%
Multi-agent failure modes (MAST)
14
Task horizon doubling time
~89 days
Registry incidents
6

Evidence

Failure taxonomies caught up with agents in 2025

Documented cases: failure Modes
DateCaseCountryWhat happenedSource
Mar 17, 2025MAST taxonomyUnited States14 failure modes across seven multi-agent frameworksSource
Jun 4, 2026Microsoft agentic failure taxonomy v2United StatesAdded MCP and plugin abuse and computer-use attacksSource
Nov 19, 2025Vectara leaderboard rebuiltUnited States and ChinaDeepSeek-V3.2 near the 3.3% best; several frontier models above 10%Source
Jan 29, 2026METR Time Horizon 1.1United StatesClaude Opus 4.5 at 320 minutes, GPT-5 at 214Source

Registry

Six registry incidents bear on AI failure modes

Registry incidents where this discipline was tested in practice. Each links to its full record.

Registry incidents linked to this page
IncidentDateDeveloperRegulator action
Uber Test Vehicle Kills Pedestrian in TempeMar 2018Uber ATGArizona suspended Uber's testing on March 26, 2018. Yavapai County declined to charge Uber in March 2019. NTSB adopted its findings on November 19, 2019. The backup operator pleaded guilty to endangerment in July 2023 and received three years of supervised probation.
Air Canada Chatbot Invents a Refund RuleNov 2022Not disclosedOn February 14, 2024, the BC Civil Resolution Tribunal found negligent misrepresentation and ordered Air Canada to pay CA$812.02.
Replit Agent Deletes the SaaStr DatabaseJul 2025ReplitNone reported.
Apollo Go Robotaxi Falls Into a Construction PitAug 2025BaiduNone found.
Waymo Robotaxis Pass Stopped School BusesSep 2025WaymoNHTSA opened preliminary evaluation PE25-013 on October 17, 2025, and wrote to Waymo again on December 3. Waymo counts 12 citations tied to the behavior.
Cursor Agent Deletes the PocketOS DatabaseApr 2026Cursor; model by AnthropicNone reported.

Rules

Taxonomies are voluntary; the EU code requires loss-of-control assessment

Rules and frameworks: AI failure modes
RuleJurisdictionWhat it requiresStatus
MASTAcademicStandard labels for multi-agent failuresNeurIPS 2025
Microsoft agentic failure taxonomy v2IndustryThreat-modeling categories for agentsJun 4, 2026
GPAI Code of PracticeEUAssess loss of control and other systemic risksFrom Aug 2, 2025

Status colors: green in force, blue voluntary standard or framework, amber adopted and phasing in, gray draft or pending, red withdrawn.

Record

Page Record

Datasets
ai_trust_evidence.csv (4 rows); registry incidents (6 linked)
Key sources
MAST taxonomy
Microsoft agentic failure taxonomy v2
Vectara leaderboard rebuilt
METR Time Horizon 1.1
Last verified
October 2026