The frameworks name the threats. ATR checks for them.
Six frameworks, among them MITRE ATLAS, OWASP, NIST AI RMF and ISO 42001, describe what can go wrong. ATR is the part you can run: a rule fires on a real agent artifact. Every rule carries mappings into all six frameworks, and CI enforces it.
Every recall below names its lane. Hunt comes first, then enforce wherever the measurement recorded it. The hunt lane is the engine default and loads every rule maturity; the enforce lane, the one to run with blocking on, loads only stable rules, so its figure is the one to read before turning blocking on. False-positive figures were measured in the hunt lane only.
Recall on the PINT-format corpus is 65.4% (hunt) / 0.4% (enforce). The corpus has 850 samples, 451 of them attacks. The hunt lane flagged 0 of 399 benign samples (ATR 4.1.5, measured 2026-10-15). Partly closed-book: some hunt-lane rules were mined from this corpus's own misses, so the hunt figure is not fully out-of-sample. ATR built this corpus itself from deepset and Lakera Gandalf data. It is not Lakera's official private PINT benchmark. Only a small subset of rules fire on it, so read it as a score for the prompt-injection family, not overall coverage.
Recall on a seeded subset of 4,780 of the 601,757 HackAPrompt competition prompts is 83.1% (hunt) / 49.0% (enforce) (ATR 4.1.5, measured 2026-10-19). The subset is partly in-sample. 20 rules name HackAPrompt in their YAML, and several were mined from this subset. Without those 20 rules the same samples score 28.7% (hunt) / 0.7% (enforce): the ablation reference, which approximates an out-of-sample figure but does not guarantee one: a rule that does not name the corpus can still have been tuned on it. The corpus is 100% adversarial with no benign samples, so it cannot give a precision or false-positive figure. Do not read it as evidence of precision.
Recall on the 341 internal self-test samples is 96.6% (hunt) / 3.4% (enforce). The hunt lane flagged 0 of 20 benign samples (ATR 4.1.5, measured 2026-10-15). This corpus is separate from the SKILL.md benchmark. It is also in-house, so it is not out-of-sample evidence.
Recall on garak's in-the-wild jailbreak set of 650 prompts is 92.3% (hunt) / 0.9% (enforce). On the full garak probe suite of 3,475 prompts it is 57.2% (hunt) / 0.5% (enforce) (ATR 4.1.5, measured 2026-10-15). Both are partly closed-book: many of the rules that fire on them were written from them, and the full suite contains the in-the-wild set twice, so neither hunt figure is an out-of-sample result.
OWASP Agentic Top 10
10/10 categories, each backed by rules that fire. Detections, not a checklist.
OWASP Agentic Skills Top 10 (AST10)
7/10 categories with rule coverage. AST07, AST08, and AST10 have no rule. The mapping lists AST07, AST08, and AST10 as process or meta-level concerns, which pattern matching cannot detect. Status counts stable (enforce-lane) rules only; it does not mean the control is met.
SAF-MCP
55/78 active SAF Framework Model v2 techniques have at least one reviewed ATR rule mapped to them, as partial or adjacent. Each pair was proposed from the rule's actual patterns and kept only when two independent reviewers agreed. These are candidate mappings. SAF has not validated them, and they make no claim about detection effectiveness. An adjacent rule adds supporting context. On its own it does not detect the technique. Techniques with no mapped rule are listed at the end of the mapping.
Full SAFE-MCP mapping on GitHubMITRE ATLAS
Every rule's YAML carries a MITRE ATLAS reference. It is part of how ATR maps each rule into six frameworks: ATLAS, OWASP Agentic, OWASP LLM, EU AI Act, NIST AI RMF and ISO 42001. A rule with no mapping does not reach main. CI enforces that. The rule explorer groups them by tactic.
Browse rules with MITRE mappingsInterests: the maintainer founded and controls PANGUARD AI, INC., which sells AI agent security products built on ATR.
