Methodology library
AI Risk Library
A public catalog of the risk categories our testing methodology is designed to measure — what each detector checks, what it needs, and where its limits are. This library describes our methodology. It is not a live testing service, and no customer systems are tested through this site today.
Pilot in development: the technical testing platform is being built and validated against synthetic systems first. Nothing here is a certification, a scan result, or a compliance determination.
Prompt injection and instruction-hierarchy attacks
- Version 0.1.0Direct prompt injectionDirect attempts to override the system prompt or instruction hierarchy through the user turn: instruction override, persona hijack, delimiter and formatting tricks. Measures whether the system keeps its assigned role and refusals under attack.Methodology reviewed: 2026-11-08
- Version 0.1.0Indirect prompt injectionInjection embedded in retrieved documents or tool output rather than the user turn — the attack surface of RAG and agent systems. Measures whether content the system reads can steer it away from its instructions.Methodology reviewed: 2026-11-08
Sensitive information leakage
- Version 0.1.0PII leakage (synthetic canaries)Synthetic personal-data canaries are placed in the system context; the detector measures whether the model reproduces them in answers where they do not belong. Canaries are fabricated markers, never real personal data.Methodology reviewed: 2026-11-08
- Version 0.1.0Secret and credential leakage (synthetic canaries)Synthetic API-key-like and password-like canaries in configuration or context; measures whether the system will reveal them on request, including requests framed as authorized administration.Methodology reviewed: 2026-11-08
Hallucination and unsupported factual assertions
- Version 0.1.0Unsupported factual claimsMeasures how often answers assert facts that have no support in the provided reference material. An LLM-as-judge estimate, not verified truth; calibration and human adjudication rules are in EVALUATION_METHODOLOGY.md.Methodology reviewed: 2026-11-08
- Version 0.1.0Citation groundingFor answers that cite sources: whether each cited source actually supports the statement it is attached to. Catches confident answers with decorative or mismatched citations.Methodology reviewed: 2026-11-08
RAG retrieval relevance and answer faithfulness
- Version 0.1.0Retrieval contextual relevanceWhether the retrieved context is relevant to the question asked. A retrieval-side metric: it isolates the search/index layer from the generation layer so failures can be attributed correctly.Methodology reviewed: 2026-11-08
- Version 0.1.0Answer faithfulness to retrieved contextWhether the generated answer is faithful to what was retrieved — a generation-side metric that complements retrieval relevance. Unfaithful answers on relevant context point at the generator, not the index.Methodology reviewed: 2026-11-08
- Version 0.1.0Context sufficiencyWhether the retrieved context contains enough information to answer the question at all. Separates 'the corpus does not know' from 'the system answered badly' — two different remediation paths.Methodology reviewed: 2026-11-08
9 detectors · 4 categories · Synthetic systems only — no live customer endpoints