The CVEs behind agents, in one list
Every CVE on the official record (CVE.org, enriched by NVD) that is about an AI agent product, an agent framework, a protocol implementation such as an MCP server or client, or the infrastructure agents call through. Severity, descriptions and affected versions are the official record's own; the category, attack surface and links to agents and traces are this registry's, and how each CVE qualifies is written down. Rebuilt every Monday; no score of ours, no exploit details, ever.
166 agent product · 521 agent framework · 359 protocol implementation · 381 infrastructure agents call · 335 agent-related (no listed product) · as of 2026-10-26. Rows are listed by the matching rules; a wrong listing is fixed with a one-line override, never by editing the record.
Loading the registry…
How a CVE gets on this list
- It names a listed product — one of the 41 agent products, frameworks, protocol implementations and runtimes this registry tracks, matched by the record's CPE or exact product name. General-purpose platforms with an AI feature are deliberately not listed, because listing a product lists every CVE it has.
- Or the record itself is about agents — it carries CWE-1427 (improper neutralization of generative-AI prompts), or its official description uses the agent vocabulary (tool calling, MCP server, prompt injection, agent framework …) in a way the rules spell out.
- A person can override either way.
agentdossier cves excluderemoves a false positive with a note;includemarks a row reviewed. Overrides live in one file in the repository, so every decision is on record and survives the weekly rebuild. The maintainer's guide.
Sources: CVE Program (CVE.org) records via the NVD API 2.0, CVSS and CWE as NVD publishes them (NVD's score first, then the CNA's), and the CISA Known Exploited Vulnerabilities catalog. A CVE in GitHub's advisory database that has no CVE id is not listed. References tagged Exploit by NVD are dropped. This list is a finding aid, not a vulnerability-management feed: confirm against the official record before acting.