Skip to content
Agent Threat Rules (ATR)中文
Responsible use

A public attack-detection standard must answer for dual use.

ATR's 836 rules describe attack patterns against AI agents so that they can be detected. A standard that documents attacks can also be misread as a recipe for them. CVE, Sigma, YARA and ATT&CK all carry the same tension. ATR's answer is transparency: the rules are public and reviewable, and the boundaries are written down here instead of relying on obscurity. This page covers the design intent, the known dual-use risks, what ATR does not catch, and how to report misuse.

Intended use

ATR is a defensive detection standard.

Integrate into security scanners

Run ATR rules in CI/CD pipelines, agent runtimes or MCP server middleware to detect known attack patterns in SKILL.md files, tool descriptions and agent configs. Cisco AI Defense's open-source skill-scanner has merged the ATR rule pack, and Microsoft AGT carries merged ATR example directories. Both are worked examples you can read.

Red-team baseline

Within an authorized scope, use ATR rules to measure how much of what your red-team tool finds already has detection coverage, and how much is new. The new part is where the next rule comes from.

Standards mapping and compliance documentation

Every ATR rule carries crosswalks to OWASP, MITRE ATLAS, NIST AI RMF, EU AI Act and ISO 42001. When a rule fires, it provides detection evidence supporting the requirements it maps to, which you can show to auditors or procurement. It does not mean the rule satisfies those requirements. ATR wrote these mappings; the framework owners have not reviewed or endorsed them.

Research and academic citation

Cite specific rule IDs (e.g. ATR-2026-00440) as executable detection baselines in attack research. IDs follow CVE/CWE-style numbering and never change after publication, so papers, CI scripts and external docs can safely cite them for the long term.

What constitutes misuse

ATR rules should not be used as attack generators.

Each ATR rule contains a regex describing an attack pattern, plus test cases covering both true_positives and true_negatives. Those test_cases exist to prove the rule detects what it claims and does not fire on benign input. They are how the standard stays conformance-testable. They are not a supply of ready-made attack payloads.

The following uses constitute misuse:

  • ▸Using true_positive test cases directly as attack payloads against production AI agent systems without authorization.
  • ▸Reverse-engineering ATR rule regexes into evasion variants and deploying them against unauthorized targets.
  • ▸Building attack automation tools grounded in the ATR rule corpus with the intent of breaching systems that have deployed ATR rules.
  • ▸Using ATR data to train attack models aimed at generating adversarial inputs that evade ATR detection.

The MIT license permits any use, including commercial use and forks. This definition of misuse is not a legal restriction. A standard cannot control how it is used, and should not pretend to. It is a clear statement of design intent that adopters can cite when they evaluate risk.

Dual-use disclosure

This standard is dual-use, and this page says so in the open.

ATR rules describe the behavioral signatures of attacks on AI agents. Every system that documents attacks carries the same tension: CVE databases, Sigma rules, YARA signatures, MITRE ATT&CK. The field's long-standing consensus is that writing techniques down in the open gives defenders a bigger lead than the convenience attackers draw from the documentation. Obscurity only delays. Disclosure lets every defender close the gap at once. ATR sits on the side of transparency, because a detection standard is only worth as much as it can be reviewed.

Two deliberate design constraints narrow that tension:

  • 1.The true_positive test cases are minimal pattern examples, not complete attack chains. They are enough to validate detection, but turning one into a working exploit still takes additional attack engineering. The signal a detection standard needs is a small fraction of the work weaponization takes.
  • 2.Rules for the most dangerous cases (high CVSS, actively exploited in the wild) follow responsible-disclosure timelines. Before the PR is filed, the affected vendor is confirmed to have had time to patch. ATR-2026-00440 and ATR-2026-00441 cover two critical Microsoft Semantic Kernel CVEs (CVE-2026-26030 and CVE-2026-25592), and they followed this rule: they were published only after MSRC's public disclosure.

The data sources have limits too. AdvBench, HarmBench and JailbreakBench are classified as test corpora (data/test-corpora/), not rule sources. Those datasets describe target behaviors, not wrapped attack payloads.

Finally, honest disclosure means stating what ATR cannot catch. The rules are pattern-based. By construction they are blind to paraphrase, semantic equivalence, non-English injection, gradual multi-turn escalation and transport-layer protocol attacks. Any attacker who reads the published rules can route around the specific verbs, nouns and syntax. ATR publishes known evasions in its rules' evasion_tests fields (fewer than half of the rules have them) and keeps a full LIMITATIONS.md, so these boundaries are a document you can cite instead of a blank in the marketing.

Abuse reporting

If you observe ATR being misused.

A standard cannot prevent misuse, and should not pretend to. The MIT license leaves no room to restrict use, and trying would not be good engineering. Misuse cases are still worth knowing about. They feed back into the documentation and the rule design, so the next revision is clearer and harder to turn around. If you see ATR being misused, tell the maintainer.

Abuse reports
[email protected]
General questions
[email protected]
Security vulnerability reports (ATR itself)
Follow the SECURITY.md process