VULNpocalypse
InformationalEcosystem changeMonitoring

Mythos-era exploitation statistics: 14 of 1,061 AI-found vulnerabilities exploited, and what Firefox 150’s 271 fixes do and do not show

VulnCheck found 1.3% of AI-found vulnerabilities exploited in 1H 2026, the same as all bugs. Patch AI-found CVEs by exposure, not by how they were found.

Published
Updated
Fixed in
Firefox 150 (271 Mythos-found bugs); Ghost 6.19.1 (CVE-2026-26980)
Exploitation
14 of 1,061 AI-attributed vulnerabilities confirmed exploited in 1H 2026 (VulnCheck); the only Anthropic-credited one named is CVE-2026-26980 in Ghost

Live scores

CVE-2026-26980
EPSS
0.711
very likely to be attacked soon
VEDAS
0.989
strong exploit evidence
On CISA KEV
Not listed
Detection rule
Nuclei · Suricata

What happened

The exploitation numbers. VulnCheck’s “State of Exploitation 1H 2026” report, published on 28 July 2026, combined two datasets: a tracker of Anthropic-credited CVEs kept by VulnCheck researcher Patrick Garrity, and the UC Berkeley Vulnerability Research Initiative. Together they attribute 1,061 vulnerabilities to AI-assisted discovery. Of those, 14 (1.3%) are confirmed exploited in the wild, and VulnCheck saw four of the 14 used against its own canary systems (VulnCheck).

VulnCheck says that rate is “roughly matching the overall exploitation rate of all vulnerabilities” in the same six months, and concludes “the data does not suggest that AI discovered vulnerabilities are inherently more likely to be exploited” (VulnCheck). Garrity added that the impact “has been real but modest, and it’s worth watching closely” (Security Boulevard).

Limits to keep in mind:

  • The window is short. Project Glasswing and Claude Mythos Preview launched in April 2026, so the half-year covers about three months of Mythos-era disclosures (Security Boulevard; Help Net Security).
  • Most AI findings have no CVE yet. Anthropic reported more than 23,000 findings, of which 126 had become published CVEs at the time of the report (VulnCheck).
  • Only one exploited example is named: CVE-2026-26980, an Anthropic-credited bug in Ghost CMS. The other 13 are not listed (VulnCheck). CVE-2026-26980 is an unauthenticated SQL injection in Ghost’s Content API, affecting 3.24.0 to before 6.19.1, rated CVSS 9.4 and fixed in 6.19.1 (CVE record).

The same report gives the baseline for all vulnerabilities. VulnCheck added 495 KEVs in 1H 2026. 23.43% showed exploitation on or before the day the CVE was published, down from 28.93% in 2025. The median time from CVE publication to KEV fell to 80 days, from 120 in 2025 (VulnCheck).

Firefox 150. Mozilla said on 21 April 2026 that Firefox 150 “includes fixes for 271 vulnerabilities” found when it ran an early version of Claude Mythos Preview against Firefox. An earlier run with Opus 4.6 led to fixes for 22 security-sensitive bugs in Firefox 148 (Mozilla). Mozilla’s headline used the phrase “zero-days”, but the body says “vulnerabilities” and “bugs” throughout, and gives no severity breakdown. Mozilla also wrote that it had not “seen any bugs that couldn’t have been found by an elite human researcher” (Mozilla). Commenters on Bruce Schneier’s post disputed the “271 zero-days” framing, and one said the Firefox 150 release notes tag only three bugs as found with Claude. We have not checked that count (Schneier on Security).

Am I affected?

Everyone who patches software is affected indirectly, because AI tools are raising the number of CVEs to triage. Two direct checks:

bash
# Firefox: 150 or later carries the 271 Mythos-found fixes
firefox --version
# macOS
/Applications/Firefox.app/Contents/MacOS/firefox --version

# Ghost CMS: CVE-2026-26980 is fixed in 6.19.1
ghost version            # run in the Ghost install directory
npm ls ghost

For the wider question, look at how your vulnerability process ranks findings. If it raises or lowers priority because a CVE was “AI-found”, that rule is not supported by the 1H 2026 data.

What to do now

  1. Update Ghost to 6.19.1 or later if you run it. CVE-2026-26980 is the named exploited AI-found bug, and it needs no login (CVE record).
  2. Update Firefox to 150 or later on every desktop, and keep auto-update on. The Mythos-found fixes are bundled into normal releases (Mozilla).
  3. Triage AI-found CVEs by exposure and evidence, not origin. Use KEV listings, vendor exploitation notes and whether the component faces the internet. VulnCheck’s data shows no extra exploitation risk from the discovery method (VulnCheck).
  4. Plan for the 80-day median, and for the 23% that are faster. Nearly a quarter of 1H 2026 KEVs were exploited on or before disclosure day. For internet-facing products, patch on release, not on KEV listing (VulnCheck).
  5. Expect the write-up, not the patch, to start the clock. The Mythos-found Rejetto HFS bug went 11 weeks without attacks, then was probed the day after the technical write-up (our HFS advisory). Watch for public proofs of concept on CVEs you have deferred.

Why this matters

The 1.3% figure is easy to misread as “AI-found bugs are safe to ignore”. It means they are exploited at the normal rate, so they need the normal response. What changes is volume. For the scale of what is still unpatched, see our Project Glasswing advisory.

Sources