Security Desk · live register

AI-Found Vulnerabilities

47 vulnerabilities found by AI tools, each one checked at its official CVE record or vendor advisory

Real vulnerabilities that an AI tool found, from Google Big Sleep and XBOW to OpenAI's Aardvark and the open-source Vulnhuntr, compiled by YFarmX from the public CVE Program records and the vendors' own advisories. Each row shows how strong the evidence is, from a CVE record whose own credits name the AI tool down to a claim that rests on the finder's own write-up, and links straight to the record.

Checked 19 Sept 2026 · 69 entries · 47 at a record

How this register is kept

what the number is, where it comes from, and how each row is checked

What the number is. 47 vulnerabilities on this register carry an identifier we checked at a primary record, a CVE record, a GitHub advisory or the merged fix, and are credited to an AI finder: 20 where the CVE record's own credits name the finder, 10 where the assigning CNA credits it in its own advisory, and 17 where the credit rests on the finder's or a researcher's publication against a record whose credits are empty or name people. Every row was checked against the CVE Program's cvelistV5 mirror on 12 and 19 September 2026.

Where it comes from. The CVE Program publishes every record for free public use under its terms of use. The register reads each record from the CVEProject/cvelistV5 mirror and takes the credit from the record's own credits field, then from the vendor's release note or advisory where the CNA files credit there, then from the finder's own publication. Each row links its record and the source that credits the AI, and its own page carries the credit verbatim.

The five tiers, strongest first. Credits array: the CVE record's own credits field names the finder, as CVE-2025-0133 does with XBOW and CVE-2024-9143 with Google OSS-Fuzz-Gen. CNA advisory: the assigning CNA credits the finder in its own release note or security page, which is how Apple and Chrome file credit. Outside claim: the finder or a researcher claims the row and the record's credits are empty or name people, which is where every Vulnhuntr, Aardvark and AISLE row sits. Awaiting record: an identifier quoted in public that the CVE Program's list will serve once the record publishes; the 9 rows here are depthfirst's FFmpeg identifiers, which answered 404 at the mirror on 12 and 19 September 2026. Aggregate total: a published headline count, such as XBOW's roughly 1,060 HackerOne reports; the 10 rows here are kept as their own tier because a total is a claim about a count rather than a record.

Reproduce a row. The weekly refresh is one command per row, run against the CVE Program's own mirror, reading the credits array rather than any announcement:

curl -sS "https://raw.githubusercontent.com/CVEProject/cvelistV5/main/cves/2026/22xxx/CVE-2026-22795.json" \
  | python3 -c "import json,sys; d=json.load(sys.stdin)['containers']['cna']; print(d.get('credits'))"

The largest block is Google Big Sleep, 26 identified rows, with its origin in the October 2024 SQLite find. The control row is a human find: CVE-2026-87491, which headlines conflated with an AI exploit chain and which Google's release note credits to a person with a $2,500 reward.

Showing 69 of 69 entries

Overview

Checked at a recordCVE record, advisory or merged fix47of 69 entries · checked 19 September 2026

Strongest tierThe CVE record's own credits name the finder2010 more credited in the CNA's own advisory

Google Big SleepIdentified rows2618 at the record, 7 in Apple's and Google's advisories

Awaiting a CVE recordQuoted in public, 404 at the mirror9depthfirst's FFmpeg identifiers, re-queried 12 and 19 September

Aggregate totalsHeadline counts, their own tier10from XBOW's roughly 1,060 reports to CodeMender's 72 patches

Published bountyAmong identified rows1of 47, and it is the human control: $2,500 for CVE-2026-87491

Entries · by finder

Google Big Sleep2637.7%
depthfirst1014.5%
Vulnhuntr1014.5%
Other programmes and reports68.7%
OpenAI Aardvark57.2%
AISLE57.2%
Google OSS-Fuzz-Gen22.9%
XBOW22.9%
OpenAI o3 (Sean Heelan)22.9%
Human researcher (control)11.4%

Entries · by month

Dated by the record's datePublic where it carries one, else the advisory or write-up date; the last twelve months of 24 tracked.

Evidence tier · all entries

69entries

Credits array: 2029.0%
CNA advisory: 1014.5%
Outside claim: 2029.0%
Awaiting record: 913.0%
Aggregate total: 1014.5%

Every entry · most recent first

Questions people ask

Which CVEs has Google Big Sleep found?

Eighteen CVE records carry Big Sleep in their credits, checked at the CVE Program's cvelistV5 mirror on 12 September 2026: CVE-2025-6965 in SQLite (credited to Vlad Stolyarov of Google's Threat Analysis Group with assistance from Big Sleep), seven FFmpeg records CVE-2025-59728 to CVE-2025-59734, seven QuickJS records CVE-2025-62490 to CVE-2025-62496, CVE-2025-9086 in curl, CVE-2025-11731 in libxslt and CVE-2026-83596 in WebKitGTK. Seven more are credited in the assigning CNA's own advisory: five Safari records (CVE-2025-43429, 43430, 43431, 43433 and 43434) on Apple's Safari 26.1 page, and two Chrome records (CVE-2025-9478 and CVE-2025-9132) in Google's release notes, all read on 19 September 2026. That is 26 identified rows, plus the October 2024 SQLite find that was fixed before it needed a CVE.

How many vulnerabilities have AI tools found?

47 on this register carry an identifier checked at a primary CVE record, GitHub advisory or vendor advisory as of 19 September 2026: 26 Google Big Sleep, 6 Vulnhuntr CVEs plus 2 pull requests, 5 OpenAI Aardvark, 4 AISLE on OpenSSL, 2 kernel bugs found with OpenAI o3, 1 Google OSS-Fuzz-Gen, 1 XBOW and 1 human-found control. The larger public figures are headline totals, kept in their own aggregate tier because a total is a claim about a count rather than a record: XBOW's roughly 1,060 HackerOne reports, AISLE's 225-plus CVEs, depthfirst's 21 FFmpeg zero-days, Anthropic's 10,000-plus Glasswing flaws and CodeMender's 72 patches.

Has the CVE Program published the nine depthfirst FFmpeg identifiers?

CVE-2026-39210 to CVE-2026-39218 answered HTTP 404 at the CVE Program's cvelistV5 mirror on 12 September and again on 19 September 2026. depthfirst's write-up, published 2 June 2026 and opened on 19 September, says nine of its 21 FFmpeg findings "have already been assigned CVEs" and lists twelve more as DFVULN-116 to DFVULN-127 with no identifier yet. The register holds the nine as claimed until a record serves, and re-queries them on each refresh.

Do AI-found vulnerabilities earn bug bounties?

One of the 47 identified rows carries a published payment, and it is the human control: $2,500 to Jihyeon Jeong for CVE-2026-87491 in Chrome, per Google's release note of 8 September 2026. Google Big Sleep, OSS-Fuzz-Gen, CodeMender and OpenAI Aardvark report upstream to projects that pay nothing, as do the kernel and OpenSSL rows. XBOW's PAN-OS finding and the Vulnhuntr set routed through paying programmes (HackerOne and huntr) whose amounts are unpublished at the record.

How strong is the evidence behind each row?

Each row carries one of five tiers, strongest first. Credits array: the CVE record's own credits field names the AI finder (20 rows). CNA advisory: the assigning CNA credits the finder in its own release note or security page, which is how Apple and Google file credit (10 rows). Outside claim: the finder or a researcher claims the row and the record's credits are empty or name people (20 rows). Awaiting record: an identifier quoted in public that the CVE Program's list will serve when the record publishes (9 rows). Aggregate total: a published headline count (10 rows). The confidence marker on each row is carried from the research note.

Each row opens the entry's own record page, with the credit as recorded and the sources; the small link under it opens the primary record. Download the data: CSV ·JSON ·RSS · CC BY 4.0 with attribution to YFarmX.

Severity is the record's own CVSS or the CNA's rating. A claim is reported as a claim, with the tier it rests on.Last checked · 19 Sept 2026