PRODUCTION SCENARIO
A support assistant runs a PII detector and a prompt-injection classifier on every request. Across a two-week staging run over replayed production traffic, neither classifier fired once. The team reads that as evidence the guardrails work and asks to launch.
Which evidence should the review require instead?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Measured false-negative rates from a purpose-built adversarial set
Replayed benign traffic contains no attacks, so zero firings measure nothing; what matters is how often a deliberately crafted injection or an unmasked identifier gets through, which only a red-team set written for this application produces. Vendor benchmarks come from a different distribution and threat model. A longer replay of the same clean traffic multiplies the same non-result. Latency headroom answers a capacity question, not a safety one.