Security rules
for AI agents.
Agents read web pages, open mail and install tools, and they follow instructions hidden inside them. ATR writes documented attacks down as rules, each with an ID and tests. The rules run in your own environment and report what they match. Anyone can use them.
Update the client to the new auth API. The docs are at docs.example.dev/auth.
Verdicts are what agent-threat-rules 4.3.0 actually returned. Each attack is taken from a rule's test cases.
- Effective rules
- 836
- Threat categories
- 10
- Organizations that merged it
- 13
- License
- MIT
- CIRCL
- OWASP Foundation
- FINOS
- Gemara project
- IBM
- Cisco
- Gen Digital
- NVIDIA
- AG2
- SigmaHQ
- Microsoft
- Tencent Zhuque Lab
A documented attack gets a name.
Ten categories, from a poisoned tool description to a goal quietly swapped for another. Each panel opens the rules in its category.
- Prompt InjectionCrafted inputs that hijack what an agent does253 rules
- Skill CompromiseMalicious or vulnerable MCP skills and SKILL.md files52 rules
- Context ExfiltrationStealing conversation context and sensitive data138 rules
- Agent ManipulationSocial engineering that talks an agent into behaving differently108 rules
- Tool PoisoningPoisoned tool descriptions and malicious tool responses116 rules
- Privilege EscalationAn agent gaining capabilities nobody authorised77 rules
- Excessive AutonomyAgents going past the limits they were meant to work within38 rules
- Data PoisoningTainted training data, memory or retrieval sources that bend what an agent does9 rules
- Model AbuseMisusing what a model can do: jailbreaks, harmful output, burning resources41 rules
- Model SecurityAttacks on the model itself, such as extraction, inversion and adversarial inputs4 rules
- MITRE ATLAS
- OWASP Agentic
- OWASP LLM
- NIST AI RMF
- EU AI Act
- ISO/IEC 42001
We used to have decades.
Forty years passed between the first powered flight and the treaty that gave the world common rules for flying. Software flaws got shared numbers only after the internet was everywhere. Each time, we could afford to wait.
Agents are different. Models turn over in weeks, and agents already work in hospitals, banks and government offices. Attacks spread fast. Defenders are still trading emails and waiting on reports.
ATR does one small thing: it writes down documented attacks so the next person can spot the same trick.
An attack one person hit becomes everyone's rule.
Take ATR-2026-02403: it reached users 20 days after it was reported. Its maturity is test, so it runs in the alert and hunt lanes and the enforce lane does not load it.
- Incident
Someone finds a new way in
July 2026: Manifold Security reports that an Azure DevOps MCP server hands pull-request descriptions to the agent untouched. An HTML comment in the description is invisible on the web page. The agent reads it and does what it says, with the reviewer's own access.
- Reported
- 2026-07-22
- Reporter
- Manifold Security
- Rule
It is written down once
The attack becomes a rule with a permanent ID. It describes the shape of the attack, not the vendor, so the same trick is caught wherever it shows up.
- ID
- ATR-2026-02403
- Rule dated
- 2026-07-28
- Tests
It is tested both ways
Every rule carries examples it must catch and examples it must leave alone. CI reruns them on every pull request and every push to main.
- Must catch
- 8
- Must leave alone
- 14
- Release
Everyone gets it
It ships in the next numbered npm release. Any agent that updates ATR gets the rule, and it runs in the lanes the rule's maturity allows. No one has to sign anything or pay for it.
- Reached main
- 2026-08-03
- Shipped on npm
- 2026-08-11 · 3.5.12
One rule, mapped to the frameworks you answer to.
- ASI01:2026
- ASI06:2026
- LLM01:2025
- LLM05:2025
- AML.T0051.001
- AML.T0057
- EU AI Act Art. 14, 15
- NIST AI RMF Map · Measure · Manage
The mappings on the left come from the rule file itself. Each crosswalk on the right is a public document in the repository. ATR wrote these crosswalks; the bodies named have not reviewed or endorsed them. Some mappings are known to be wrong and are being redone. Until they are fixed, use them as a reference only, not as audit evidence. See the errata
Spot the attack, pick the tests, prove it.
- ATR RulesReleased
Spot the attack.
Rules that run inside your own process. Nothing has to leave.
- ATR RouterNot yet public
Work out what to test.
Reads what your agent can reach and keeps only the tests that matter for it.
- ATR EvalSource public · no release yet
Then prove it.
Each attack runs many times. The result is signed, and anyone can recompute it.
Where you start depends on who you are.
You build agents
Your agent spends its day reading things other people wrote. Web pages, tool descriptions, files from strangers.
Run ATR on what your agent reads and on every tool call. The engine and the Claude Code hook only report. They block nothing until you turn blocking on. Some adapters are the exception: Mastra, OpenShell and NemoClaw block high and critical matches out of the box, and you can change that in their settings. Blocking is meant for the enforce lane. That lane loads only the stable rules, so it catches far fewer attacks. ATR runs inside your own process, so the data stays where it is.
You run security
Agent incidents land in the same queue as everything else, with no name and no query that finds them.
Turn the rules into the queries your SIEM already runs, and put the rule ID on the ticket. Whoever picks it up next knows exactly what it was.
You ship an AI product
Customers ask what you can stop. A feature page can't answer that. A list of rule IDs can.
Ship the rules, pass the public conformance tests, then list yourself in ADOPTERS.md with a link anyone can open.
You have to answer for it
Auditors, regulators and insurers won't take your word for it. They want something they can check themselves.
The rules run fully offline and map to the EU AI Act, NIST AI RMF and ISO/IEC 42001. ATR Eval, public as source but not yet released, is meant to measure how a given agent release holds up against hijacking.
You find attacks
You found a hole. It shouldn't get fixed only in the one system where you found it.
Send it as a rule with tests, or report it and let a maintainer write it. Bypasses and false positives count too. Anything accepted carries your name.
The rules are public and free to use. No rule is accepted for payment.
Specification 1.0.1 · DOI 10.5281/zenodo.19178002 · MIT
