CVE Tools
Google GTIG Measured the AI Bug Boom: 10,740 Disclosures a Month, AI-Found Flaws Skew RCE — and Only 1 in 431 Gets Exploited. Our own AI-stack KEV ledger hasn't added a single entry since September

Google GTIG Measured the AI Bug Boom: 10,740 Disclosures a Month, AI-Found Flaws Skew RCE — and Only 1 in 431 Gets Exploited

Our own AI-stack KEV ledger hasn't added a single entry since September 21. That's not a tracking gap — it's the same concentration GTIG's global numbers point to: the disclosure flood is real, but confirmed exploitation stays rare and narrow.

On September 30, Google's Threat Intelligence Group (GTIG) published the first large-scale dataset trying to settle an argument the industry has mostly had with anecdotes: is AI actually accelerating vulnerability discovery, and is that acceleration translating into more real attacks? Their headline numbers: monthly CVE disclosures roughly doubled from 5,045 in January 2026 to 10,740 in August, while only about 1 in 431 disclosed vulnerabilities in 2026 — 0.23% — was ever observed exploited in the wild. We maintain our own live ledger of one specific slice of this: AI/ML-stack CVEs that make it onto CISA's Known Exploited Vulnerabilities catalog. Six weeks after we last counted it, that ledger hasn't moved. Here's the delta, the gap in GTIG's own count we can't close for them, and what their report deliberately doesn't tell you.

10,740vulnerability disclosures in August 2026up from 5,045/month in January — GTIG's own count
1 in 431disclosed 2026 CVEs ever exploited0.23%, per GTIG
14AI-stack CVEs ever added to CISA KEVunchanged since our Sept 21 ledger
0new AI-stack KEV entries since then6 weeks, zero movement

The ledger didn't move

On September 21 we published the full list of every AI-stack CVE ever added to CISA's KEV catalog: 14 total, six of them Langflow, three LiteLLM, one each for n8n, MLflow, Ray, Marimo and Starlette. We re-ran that count for this piece, product by product. It is still 14. No AI-stack CVE has entered KEV since July 17, 2026 — the date Langflow's CVE-2026-9198 (auto-login bypass chained into code-validation RCE) was listed. That's not a gap in our tracking; we checked all seven products individually against our own graph. It lines up with GTIG's own finding: across the entire AI stack through August 2026, they report observing no zero-day exploitation — only n-day weaponization of bugs that were already public, patched, and in most cases already sitting on KEV before being abused further.

CVEProductKEV addedDays after publication
CVE-2026-59822LiteLLM (MCP OAuth2 bypass)2026-09-0256
CVE-2026-48710Starlette (used in a LiteLLM auth-check bypass)2026-09-0299
CVE-2026-64849MLflow (SSRF via webhook redirect)2026-08-192

Those three are the most recent entries on the ledger, and all three predate our September count. Nothing has been added since.

The case study GTIG leads with

GTIG's centerpiece example of AI cutting both ways is CVE-2026-1731, a pre-authentication OS command injection in BeyondTrust Remote Support and Privileged Remote Access. GTIG says it was found autonomously by the third-party research agent Hacktron AI.

In our own data: CVSS 9.8 (AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H), CWE-78, an EPSS score in the 99.8th percentile, and a public proof-of-concept on GitHub. CISA added it to KEV on February 13, 2026 — seven days after the February 6 disclosure.

CVE-2026-1731, by GTIG's own clock

  1. Disclosed and patched
    Pre-auth OS command injection in BeyondTrust RS/PRA, CVSS 9.8. GTIG credits discovery to the third-party agent Hacktron AI.
  2. First threat cluster exploiting it
    Within 4 days of disclosure, per GTIG.
  3. Added to CISA KEV
    Also inside GTIG's 7-day window for exploitation reaching six distinct threat clusters.
  4. Five more clusters join
    Within 7 days of disclosure, deploying SNOWLIGHT, SPARKRAT and cryptominers, per GTIG.

What skews when AI does the finding

GTIG's risk-rating comparison (its own scale, explicitly not CVSS) shows AI-discovered vulnerabilities skewing toward Medium risk (58% vs. 28% for everything else) with RCE as the leading consequence (50% vs. 26%). Their reading: AI agents are being pointed at privilege boundaries and critical infrastructure deliberately, rather than doing broad low-value scanning. Across the whole AI/ML software layer — orchestration frameworks, agent platforms, inference servers, vector databases — GTIG counts 2,076 cumulative CVEs since January 2025, more than 1,500 of them in 2026 alone. Orchestration and agent frameworks (Langflow, LangChain, Dify, CrewAI, MCP implementations and similar) account for 782 of this year's total on their own, a +347% surge — half of every AI-related flaw GTIG counted in 2026 sits in that one layer, which is also where five of our six Langflow KEV entries and all three LiteLLM ones live.

What GTIG did not release

The report vs. the gaps in it

What it gives you
  • Percentages for AI-discovered vs. not (risk mix, RCE share)
  • Aggregate disclosure and exploitation trend lines, indexed to Jan 2025
  • A handful of named examples (CVE-2026-1731, CVE-2026-42271, CVE-2025-3248, CVE-2026-5027)
  • Two named discovery tools: Hacktron AI and AISLE
  • A named remedy: CodeMender, integrated into Google AI Threat Defense
What it withholds
  • An absolute count of AI-discovered CVEs — percentages only, no denominator
  • The CVSS or EPSS of the AI-discovered set as a group
  • Named threat actors — exploitation is described only as unnumbered "threat clusters"
  • The verified lab/vendor ledgers and attribution heuristics used to tag a CVE as AI-discovered
  • Confidence bounds on any of the above
GTIG itself says public CVE data likely undercounts AI-discovered flaws, since CVE repositories carry no standardized AI-attribution field and some fixes ship silently.

Two disclosures worth sitting with

The one finding this supports

Disclosure and exploitation have risen in near lockstep this year — GTIG's indexed figures put overall CVE disclosure growth at +128% and in-the-wild exploitation growth at +127% since January 2025. The flood is real, but in aggregate it has not outrun defenders. What it has done is concentrate: 62% of everything exploited in 2026 was a zero-day, edge and security-appliance software punches far above its CVE-count weight in exploitation share, and in the one architectural layer we can check against our own graph — AI/ML middleware — confirmed KEV exploitation hasn't grown in six weeks. If your patch queue is sorted by raw CVE count or CVSS alone, this is the argument for re-sorting it by KEV and EPSS instead; our live trend data tracks that shift as it happens, not just in retrospective reports.

Method note

Every AI-stack row here comes from cve.tools' own merged CVE database (CVE.org, NVD, GHSA and BDU records reconciled into one record per CVE) matched against CISA's KEV catalog by product name — the same method behind our September 21 ledger, re-run live for this piece. GTIG's percentages come from their published report; they do not share the underlying CVE list or raw counts, so we can verify their named examples against our own data but cannot independently reproduce their aggregate statistics. CVSS is NVD's score where one exists; EPSS is today's value, not the value at time of exploitation.

CVE, KEV and EPSS data verified live; GTIG figures as published 2026-09-30 2026-11-05

Sources: GTIG — Vulnerability Discovery and Exploitation Trends in the AI Era, Microsoft — Insights from the 2026 Microsoft Digital Defense Report, CISA Known Exploited Vulnerabilities Catalog, CVE.org record for CVE-2026-1731.