AI Science Tracker

AI-discovered vulnerabilities

Public disclosures, zero-day context and documented fixes. Only findings with an original source establishing AI involvement are included.

Browse cases · Explore charts

Reading the evidence

How to read an AI vulnerability discovery

A case connects a publicly disclosed flaw to a source describing AI involvement. Open Source-attributed AI, human and vendor contributions on any case to see what the system did, how researchers used its output and what the cited sources establish.

  • Follow the evidence: Original disclosure describes the finding; AI attribution documents the system's role; Advisory / fix evidence supplies the remediation context. These links can point to the same report when it covers all three.
  • Read the dates separately: disclosure is the public report date, AI attribution is the dated account of AI involvement, and sources checked is this catalog's review date. None establishes when the flaw was introduced or first exploited. Unknown attribution dates remain explicit.
  • Distinguish cases from identifiers: a case may group multiple CVEs, and some findings have no established CVE here. Case counts and unique CVE counts therefore measure different things.
  • Check the remediation note: a documented fix is evidence about a cited correction or release. Use the vendor advisory to assess affected versions and upgrade guidance for your own installation.

Explore cases by disclosure year, discovery context or product, or search for a CVE, AI system or finding. This selected collection cannot establish which AI system discovers the most vulnerabilities: source coverage, access and reporting practices differ.

Browse tracked disclosures · Read the label definitions · Download cases (CSV) · Source records (JSON)

Catalog charts

Explore six charts and filter cases

Click a chart mark or table count to filter the list below; click again to clear that selection. Filters combine and can be shared through the page URL. Chart counts describe the full selected catalog and stay fixed while filters narrow the list. They measure published coverage, not system effectiveness or product risk.

Tracked disclosures

Loading disclosures…

Loading…

What the labels mean

This is a manually selected collection of public cases, separate from the scientific-discovery count. Each record covers one identifiable finding; aggregate model claims and benchmarks do not become individual vulnerabilities.

Undisclosed at discovery (source report) preserves a source's account of a flaw unknown to defenders at that time. It provides zero-day context, not independent proof of exploitation. Pre-release finding means the cited source says it was caught before an official release. Other disclosures are not automatically classified as zero-days.

Fix documented means a cited source records a correction or fixed release. It does not show which users installed it, whether a product is currently safe, or whether an attack occurred. Disclosure, AI-attribution and source-check dates are shown separately.

A missing CVE means no identifier is established by the cited evidence here. Findings remain attributed to the named researchers and AI systems; source checking does not independently reproduce the result. This tracker is not a live vulnerability inventory.

Download the source records (JSON) · Catalog methodology