running-prompt-injection-eval

To run a prompt-injection evaluation, supply a JSONL corpus and target endpoint to execute a structured eval harness. It validates pre-flight checks, logs attempts, classifies outcomes, and aggregates results for reporting.

Four-outcome classification in LLM red-team testing categorizes each response as blocked, passed, partial, or inconclusive. This classifies guardrail effectiveness and tracks false negatives during prompt-injection evaluations.

Yes, signed Rules of Engagement validation is required to test prompt-injection guardrails. The evaluation harness enforces scope control, corpus schema checks, and safety reviews before executing tests against the target endpoint.

Prepare a JSONL corpus for LLM guardrail audits by formatting test prompts according to required schema checks. The harness ingests the corpus, runs dry-run validations, and deduplicates findings for regression testing.

False-negative tracking during pre-deployment safety reviews reveals guardrail failures where malicious prompt injections bypassed defenses. It produces per-class summaries and finding packs to identify vulnerabilities before deployment.

Yes, you can supply existing user-supplied corpora for structured logging and eval classification. The harness validates the corpus schema, runs dispatch workflows, and generates per-class aggregation tables with deduplicated findings.