Agent Reliability Engineering (ARE)
The ARE Incident Database.
34 catastrophic agent failures that actually happened, each one cited, and indexed onto the OWASP Agentic Security Initiative (ASI) Top 10 where a category fits.
AREDB is independent. It is not an OWASP project and is not affiliated with or endorsed by OWASP. It indexes onto the ASI Top 10 because that is the vocabulary the field is standardizing on.
Every entry carries a public source, a permanent identifier, and an OWASP ASI category where one fits. Vendors, including the maintainer, may claim coverage of an entry. Every claim declares how it is verified, and claims that stop holding are withdrawn, not reworded.
Coverage claims.
A vendor may claim that its product stops one of these failures. The registry does not rank or endorse vendors. It records what was claimed, by whom, and how you can check it yourself.
- Every claim states how it is verified. Some ship a snippet on the entry page that runs from a plain package install. Others are verified by the vendor against its own component and publish no snippet. The entry says which, so a claim never implies it reproduces standalone when it does not.
- The install-only checks run on every push. 11 of the 25 claims AgentX has filed, executed by a test in this repository. The other 14 are verified by the vendor against its own component and are not re-run here.
- A claim that stops passing is withdrawn. The entry is reclassified and its identifier is kept, so the record of the claim survives.
This index is not reserved for the maintainer.
Any vendor may claim coverage, held to the same bar AgentX Core is held to here. Submit a claim, and state which of the two verification levels it carries, through CONTRIBUTING.md.
How these claims are checked.
Every coverage flag above is a claim, and a claim nobody can check is a press release. 11 of those claims ship a runnable snippet. A test in the repository scrapes the snippet out of each entry that publishes one, runs it in a clean process with no key and no gateway, and asserts two things: that the block fired, and that the tool body never executed. A block that prints a warning while the action still happens is not a block.
Those run on every push and on a weekly schedule, so a future release cannot quietly break a claim that carries one. When a claim stops being true the entry is reclassified, not reworded, and that rule lives in the governance model rather than in a promise. The badge below is the live one. If it ever goes red, it goes red here.
11 published checks, executed on every pushRead the test or read the rule that says a failing claim gets withdrawn, not reworded.
Cite it.
Incidents carry a stable identifier, ARE-2026-001 through ARE-2026-034. The identifier is permanent: an entry that turns out to be wrong is marked, never deleted and never renumbered, so a citation always resolves. The data is CC BY 4.0 and the tooling is MIT.
Use the identifiers as the shared reference for the field, the way a CVE number works. That is the whole point of a registry.
How this maps to the standards.
Several standards now cover AI agents. If your organization follows one, here is how this registry connects to it.
- OWASP Agentic Top 10. The taxonomy used here. Every incident in this registry is indexed to it. Published in December 2025 by the OWASP Agentic Security Initiative, a non-profit foundation. Free to read.
- AIUC-1. 51 requirements across six domains, revised quarterly. Published by an AI insurance company that also sells audits against it; independent firms certify to it. Cross-referenced to MITRE ATLAS, NIST AI RMF, ISO 42001 and the EU AI Act. Free to read.
- Google SAIF. Google’s Secure AI Framework. A free self-assessment, thirteen questions.
Scope of this summary. Read in full: AIUC-1 (all six domains), the SAIF self-assessment, the OWASP Agentic Top 10 titles, and two published vendor questionnaires. Not read: the HECVAT 4 AI section, the CSA AI Controls Matrix and its AI-CAIQ, and the 2026 SIG AI domain, which are spreadsheet downloads or licensed. Corrections through CONTRIBUTING.md.
Read the registry.
The full database, the taxonomy, the governance model, and the test that runs every published check all live in the open.