What has already gone wrong with AI systems, and what each case teaches
Documented incidents from courts, regulators, vendors, security researchers and the press, from 2021 to today. Each one is tagged with the layer it hits (chosen model, application and agent, operations, people), the type of source, and the SoAITrust test scenario that covers the pattern. We publish patterns and lessons, never attack playbooks.
Loading cases…
How to read the source labels. Official: court decision, regulator or vendor statement. Security research: published by the researchers who found it, usually with vendor confirmation. Press: reported by a news outlet, numbers may vary. Secondary: aggregators or vendor blogs, confirm against the primary advisory before quoting. SoAITrust observation: reported to us directly; organization not named. Ongoing lawsuits are described without conclusions on the merits.
Test your agent against these patterns
Each case maps to a scenario with an objective checker. See how agent red teaming works.
Know where your system stands
Ten minutes, questions adapted to your profile, result by email. Free pre-assessment.
Send us a case
Saw an incident in your sector? Write to [email protected]. We add it with source and without naming the organization if you ask.