OWASP LLM Top 10 coverage
A per-tenant attribution view that maps every OWASP-LLM-Top-10 category we probe to (a) the closing defense family, (b) the per-tenant state, and (c) our internal brittleness baseline. Surfaced at /coverage/llm and via GET /v1/coverage/llm.
Key points
- Five categories ship today: LLM01 (jailbreak), LLM02 (sensitive info), LLM06 (excessive agency), LLM07 (system-prompt leak), LLM10 (unbounded).
- Each category is attributed to the defense family that closes it (refusal whitelist, rate limit + PII scrub, tool-registry auth, token cap + allowlist).
- LLM01 and LLM07 drop block-rate -50pp and -100pp under paraphrasing — those are the brittleness signals we surface most aggressively.
- LLM02, LLM06, LLM10 are wording-agnostic (control-level), so paraphrasing does not change the score.
- Brittleness numbers are stamped from our internal brittleness baseline; the UI never re-derives them.
01
Per-category mapping (brittleness baseline)
- LLM01 (jailbreak) → refusal-phrase whitelist · originals 67% · paraphrases 17% · brittleness -50pp.
- LLM02 (sensitive info) → rate limit + PII scrub · originals 0% · paraphrases 0% · brittleness N/A.
- LLM06 (excessive agency) → tool-registry auth · originals 0% · paraphrases 0% · brittleness N/A.
- LLM07 (system-prompt leak) → refusal-phrase whitelist · originals 100% · paraphrases 0% · brittleness -100pp.
- LLM10 (unbounded consumption) → token cap + allowlist · originals 0% · paraphrases 0% · brittleness N/A.
02
Per-tenant state machine
- open — at least one open finding tagged with the category.
- remediated — had findings, all now closed.
- tested — at least one completed run probed the category with no finding.
- untested — no completed run has probed the category yet.
03
What this view is NOT
- Not real-time: brittleness numbers update only when we revise the baseline corpus.
- Not a substitute for the per-finding detail view at /findings/{id}.
- Not a guarantee: five categories sampled against 17 probes cannot speak to the corpus of attacks in the wild.