Submit incident

AI safety failure incidents

Definition

Safety failure incidents document cases where an AI system caused, contributed to, or failed to prevent physical harm, death, or serious safety risk to people, infrastructure, or critical systems.

Included

  • Autonomous vehicle accidents involving AI driving or perception systems
  • Medical AI errors leading to incorrect diagnosis, treatment, or dosage recommendations
  • AI-controlled industrial or robotic systems causing physical harm to workers
  • AI failures in safety-critical infrastructure such as aviation, energy, water, or transport
  • AI systems that failed to detect or escalate dangerous conditions as designed

Excluded

  • Near-misses without documented harm or credible imminent risk
  • Incidents where AI played no direct safety-critical role
  • Theoretical safety concerns or future risk scenarios without a documented event
Safety failure incidents over time
Documented incidents per year of occurrence. *2026 to date.
Timeline chart: requires category field in dataset (planned)
Unit: incidents · Source: AI Incident Index v2026.09 · Counts reflect documented incidents, not prevalence

Incident records12

Jun 2026Researchers Tested OpenAI's GPT Store. Most of It Failed.OECDJul 2025California's DMV Halts Tesla Over What 'Full Self-Driving' Actually MeansOECDJun 2025The Bomb Case That Should Worry AI Companies More Than the FBIAIIDOct 2022Russia and Ukraine Just Showed What Autonomous Naval Warfare Looks LikeOECDJun 2022One Number From NHTSA Should Worry Every Driver-Assist BuyerOECDApr 2022A Cruise Robotaxi Held Its Ground While a Fire Truck WaitedAIIDJan 2021A Braking Glitch in Tesla's Self-Driving Beta Forced a Recall of 11,700 CarsAIAAICAug 2020Guangzhou Puts Driverless Cars on Public Roads Before the Rules Catch UpOECDJul 2019The Verdict Cleared Tesla. It Didn't Answer the Real Question.AIIDAug 2017A Tesla Clipped a Curb in Redmond. The Suspension Wasn't the Real Damage.AIIDJul 2015A Factory Robot Killed a Worker Because Nobody Had to Prove It Was Safe to ApproachAIIDJan 2015144 Deaths, One Database Nobody TrustsAIAAIC
Browse all incidents →