• METR's August 14, 2026 note found CVE disclosures in cURL, OpenSSL, Firefox and Microsoft products grew far faster in 2026 than in 2025, with a share explicitly marked as AI-discovered.
• Confirmed-exploited vulnerabilities in CISA's KEV catalog grew only about 10% in the first half of 2026 versus the prior six months, against roughly 45% growth in raw CVE counts over the same window.
• AI-marked CVEs skewed toward lower-severity bugs — 12 of cURL's 15 AI-tagged findings were low-severity — pointing to a triage-volume problem more than a proven exploitation crisis.
• The same before/after methodology applied to AI's ability to speed its own algorithmic research found no measurable acceleration across seven established benchmarks.
• METR's own authors caution that AI-credentialing tags on CVEs are self-reported and likely undercount true AI involvement, since labs may not disclose all AI-assisted findings.

A new METR research note is the closest thing yet to a controlled measurement of where large language models are actually moving the needle on real-world work — and the sharpest signal it found sits squarely in security. Vulnerability disclosures in cURL, OpenSSL, Firefox and Microsoft's own products accelerated sharply through the first half of 2026, tracking a rise the nonprofit ties in part to AI-assisted bug-hunting. The same methodology, applied to AI's ability to speed up its own algorithmic research, found nothing: across seven established optimization benchmarks, METR could detect no measurable acceleration at all.

The vulnerability numbers

METR's August 14, 2026 note, authored by Tom Cunningham and Nate Rush, tracked CVE counts across four codebases through mid-2026. cURL rose from 9 disclosures in all of 2025 to 36 by June 24, 2026, with 15 explicitly AI-marked. OpenSSL went from 6 to 39, with 18 corroborated as AI discoveries. Firefox climbed from 210 to 342, and Microsoft's security-update CVEs rose from 1,243 to 1,927 — an annualized rate roughly 2.5 times 2025's, though only 1.3% carried explicit AI markers.

Discovery is outrunning confirmed exploitation

The more consequential finding sits in what METR treats as a lagging indicator. CISA's Known Exploited Vulnerabilities catalog, which lists flaws confirmed as actively exploited rather than merely disclosed, grew only about 10% in the first half of 2026 versus the prior six months, per Vulncheck data cited in the note, while raw CVE counts grew roughly 45% over the same window. METR's authors call this a significant drop in the KEV-to-CVE ratio, and caution that exploitation records can lag discovery by months or years, so the gap may narrow later.

Where the acceleration disappears

Applying the same before/after comparison to AI's ability to accelerate its own field produced a null result. Across seven benchmarks — the nanoGPT and CIFAR-10 training speedruns, the Hutter compression prize, Gurobi and MIPLIB integer-programming frontiers, Stockfish's Elo curve, and the matrix-multiplication exponent — METR found no sign of acceleration in six and only marginal, disputed movement in the seventh. Stockfish gained roughly 50 Elo per year on average since 2013, a pace one AI-credited commit in July 2026 did not visibly change.

Why the gap matters for security teams

The practical read for defenders is a volume problem before it is an exploitation problem. AI-marked findings skewed toward lower-severity, easier-to-automate bug classes — 12 of cURL's 15 AI-tagged CVEs were low-severity — meaning security teams face a larger, noisier disclosure stream to triage even where confirmed-exploited counts have not moved proportionally. METR's own caveats matter here: AI-credentialing tags are self-reported and likely undercounted, given labs' disclosed and undisclosed use of models for bug-hunting, so the true AI share of 2026's CVE growth is a floor, not a ceiling.

The verdict

METR's note is a rare data-first check on a claim usually argued from anecdote, and on security it lands closer to real but overstated as a crisis than to either nothing changed or AI is autonomously hacking the internet. The confirmed shift is in disclosure volume, concentrated in lower-severity bugs, with AI credentialing likely undercounted; the shift in actual exploitation, measured via KEV, is far smaller and lagged. Security teams should treat this as a mandate to scale triage capacity for a noisier CVE stream, not as evidence AI has yet changed attacker capability at scale.