OWASP LLM Top 10 coverage
This view presents the current product catalogue: 32 probes across all ten categories (17 locked-base + 15 expansion; commit 8752ede). Per-tenant states come only from conclusive, persisted evidence. The separate v0.5.3 locked-corpus calibration is historical methodology context, not a measure of the current engine.
Coverage maps the current probe catalogue onto the OWASP LLM Top 10 and states, per tenant, which categories hold conclusive evidence and which remain untested.
Two timelines matter and must not be conflated: the archived locked-corpus calibration used for methodology research, and the current catalogue serving tenants today. This page names both and says what the coverage view is not.
Key points
- The current product corpus includes 32 probes across all ten OWASP LLM categories, LLM01 through LLM10 (corpus expansion commit 8752ede).
- Per-tenant states and aggregate coverage appear only when completed-run evidence is conclusive and its persisted lineage can be reproduced.
- The archived methodology reference covers a locked 17-probe corpus across 5 of 10 categories; it does not measure today's engine or a tenant's defenses.
- Its ten executions per scenario (40 total; metrics record 95ac0462) repeated that same corpus rather than sampling independently; current accuracy is N/A.
- The separate mutation record (57f15b67) contains 17 originals and 85 variants across four synthetic lattice targets. No historical rate is carried into the current product.
Archived locked-corpus scope (v0.5.3 · 17 probes · 5/10)
- LLM01 (prompt injection) — included in the locked historical reference; current-engine performance in this reference: N/A.
- LLM02 (sensitive information disclosure) — included in the locked historical reference; current-engine performance in this reference: N/A.
- LLM06 (excessive agency) — included in the locked historical reference; current-engine performance in this reference: N/A.
- LLM07 (system prompt leakage) — included in the locked historical reference; current-engine performance in this reference: N/A.
- LLM10 (unbounded consumption) — included in the locked historical reference; current-engine performance in this reference: N/A.
The archived methodology reference covers a locked 17-probe corpus across 5 of 10 categories, so published methodology work stays reproducible. It is a frozen artifact: useful for reading the research, not for judging the current engine.
Per-tenant state machine
- open — at least one open finding from decision-grade evidence is tagged with the category.
- remediated — decision-grade evidence shows prior findings and all are closed.
- tested — conclusive evidence tested the category without an open finding.
- untested — the category is in the current catalogue, but no conclusive evidence has tested it.
Tenant states derive only from conclusive, persisted evidence in the current catalogue. A category the engine can probe but the tenant has not run stays untested — absence of evidence is recorded, never smoothed over.
What this view is NOT
- The historical 5/10 reference is not an accuracy estimate for the current engine or a tenant.
- Ten repeated executions of one locked corpus do not support an independent-sample confidence interval or wider extrapolation.
- Current coverage is evidence within the tested corpus, not an audit, certification, or guarantee against attacks in the wild.
The view is not an audit, a certification, or a guarantee against unseen attacks. It states where authorized evidence exists; anything beyond that boundary is a claim this page refuses to make.
Related resources