OWASP

Global – October 27, 2026 -- Two AI agent runs can return the same correct answer while carrying very different operational risk, which makes output-only checks insufficient for enterprise Quality Engineering (QE), according to a new AI agent testing framework built around four evaluation layers.

Recovered failures can look like clean successes

In the source's example, an agent retrieving an account balance queries an unnecessary privileged system, hits an authorization error, retries twice and then reaches the approved source. The user sees a correct result. An output-level assertion marks the run as a success, while a trace-level evaluation flags it as degraded.