CYBRARY Guardrail Regression
The team keeps adding guardrails and keeps breaking the product. Build a regression harness that measures both: attacks blocked and legitimate work still possible. Tune until both curves are acceptable.
Release gate RG-7 — 'guardrail change blocked 22% of valid requests'
AI engineer
QA lead
Security engineer
55 minutes · advanced
Execution trace
00:00:00policyrelease-pipeline
attack step
Guardrail v9 merged
Blocked 22% of valid support prompts in production before rollback.
00:01:00policyrelease-pipeline
attack step
No benign suite in the gate
The pipeline measured attack blocking only.
00:02:00modelassistant@v9
Legitimate refund question refused
The classifier read a normal request as an injection attempt.
00:04:00alertproduct
Rollback requested
Security and product now distrust each other's numbers.