ZDI — 47 zero-days in Meta ML libraries · Nov 11, 2026
Discovered by human researcher Mat Powell (TrendAI ZDI) and reported to Meta in March 2026. Flaws in AI libraries, not found by AI. Out under the AI-is-the-discoverer rule.
This is the other side of the ARTEX tracker. There, AI carries out attacks. Here, AI is the discoverer — finding exploitable flaws in real software before attackers do. In 2026 the finds stopped arriving one at a time: one program disclosed 5,674 in a single week, and one model produced 948 working kernel exploits in 12 hours.
Across 30 discovery events, Jan 1 – Nov 15, 2026. Dated by first public disclosure.
Conservative floor ≈ 7,482. Two large program aggregates are not itemized publicly, so up to ~32 individually-listed CVEs may already be inside them. See “Reading the total” below.
This is a cumulative bar chart — a running total, not per-week counts. Each bar adds that week’s new flaws to everything before it, so the bars only rise or stay the same height and end at the full-year total of 7,613. Weeks are Monday-start, Jan 5 to Nov 15, 2026, including zero weeks. Switch to a Per week view to see individual weeks on their own. Hover / tap a bar for details.
Same 7,613 flaws, grouped by disclosure month. October dominates because both big aggregates were announced then.
* November is through Nov 15 — GPT-6 Cyber Linux (948, Nov 4), VicOne robotics (28, Nov 1) and N-able Able2Shell (30, Nov 8), plus Fortinet Mythos (18, Nov 10), Microsoft MDASH (33, Nov 11), Forescout (66, Nov 12) and Anthropic Apache CVE-2026-119618 (1, Nov 11), grouped by disclosure date.
The 500+ aggregate may already include the Firefox CVEs (28) and FreeBSD CVE-2026-4747 (1). The 5,674 aggregate may already include Anthropic's two curl CVEs from Oct 14 (2) and the Mythos Linux CVE-2026-98432 (1). If all of those overlaps are real, the true total is ≈ 7,482, not 7,613. The Microsoft MDASH (33) and Forescout (66) finds may also already sit inside the 5,674 OSS Scanner aggregate — the kernel and embedded stacks overlap its scope — so the floor allows for those too. Anthropic has not published itemized lists, so it cannot be resolved from public sources. Individually-identified CVEs are counted separately here and flagged in place.
Unverified counts: two of the newest additions — VicOne's 28 robotics flaws and the N-able Able2Shell 30 — are counted as reported but have not been independently verified. Both counts may be revised in a later update.
Dates are first public disclosure (CVE publication, vendor advisory, or researcher writeup). Campaign aggregates are dated at the disclosure of the aggregate and flagged — their flaws were found across the window stated in each entry.
Discovered by human researcher Mat Powell (TrendAI ZDI) and reported to Meta in March 2026. Flaws in AI libraries, not found by AI. Out under the AI-is-the-discoverer rule.
ARTEX and other agentic attack campaigns are attacks, not discoveries. They live on the ARTEX attack tracker and are never double-counted here.
OSS Scanner ~23,000 unreviewed candidates; 4,824 direct-to-maintainer disclosures the dashboard itself warns may contain false positives; Dinkin & Kraft's ~100 estimated exploitable driver flaws (only the 15 confirmed are counted). Glasswing's “129,000 vulnerabilities found by AI” claim — no methodology, not counted.
17 CVEs in AI agent frameworks (Oct 2026) are vulnerabilities in AI tools discovered by researchers — the direction is reversed, so they are out.
Benchmark scores with no real-world find, CVE-Genie reproducing known CVEs, sandbox escapes and other incidents the AI caused rather than discovered — all out. Real but uncounted finds (Mythos Safari / Android / Linux chains with no CVE counts published) are noted in the baseline log.