What this video shows
The speakers group 1,368 records from a public incident database into technical strength, deployment controls, and business impact, with 13 sub-areas that they say cover about 98% of the sample. They report that weak controls such as monitoring and human review are closely associated with harm.
Use the taxonomy to ask what your current suite omits. Rank cases by affected people, reversibility, money, legal duty, and scale. Seed known failures into a safe environment, verify that the grader catches them, and require evidence for escalation and recovery paths.
Use NIST's AI Risk Management Framework to connect test priorities with organizational risk.
What you will learn
- Test deployment controls and business consequences alongside technical output.
- Spend more evaluation effort where a wrong action is costly or hard to reverse.
- Seed failures to verify that the evaluator can detect them.
- Use incidents to add regression cases rather than only to update a risk register.
How to apply this safely
- List technical, control, and business failures for one use case.
- Score impact, exposure, reversibility, and detection delay.
- Plant one safe example of each high-priority failure in the test environment.
- Confirm the grader, monitor, escalation, and recovery path all respond as intended.
Important limitations
- Public incident databases can undercount private failures and overrepresent events that attract reporting.
- A taxonomy that covers the reviewed records does not prove it covers future incidents.
- Association between weak controls and harm does not establish one control as the cause.
- New York Life employs the speakers, and the talk does not provide an independent replication package.
Sources to check
- NIST AI Risk Management Framework A voluntary framework for managing AI risks across design, development, use, and evaluation.
- 1,368 AI Failures and What They Teach About Evals AI Engineer video published October 22, 2026.
Continue learning on Learnetto
Build an AI agent test set
Create seeded failures and explicit expected outcomes.
AI evals guide
Combine code, model, and human graders.
Backtest AI agents
Replay decisions using only evidence available at the time.