Monthly vulnerability disclosures doubled between January and August, peaking at 10,740 in August, according to Google's Threat Intelligence Group. The total for the first eight months of the year has already passed the full-year count for 2025.
GTIG researchers point to a mix of automation and AI tooling for both the volume and the speed at which known bugs are being weaponized. The team noted that AI is measurably shifting how quickly vulnerabilities surface, how attackers weaponize them, and what kinds of flaws tend to draw attention.
The shape of the surge
| Metric | 2025 | Jan–Aug 2026 |
|---|---|---|
| Distinct exploited CVEs | 127 | 141 |
| Monthly average exploited | 10.5 | ~18 |
| High-risk exploited CVEs | 28 | 75 |
| Zero-day avg/month | 8 | 11 |
The numbers GTIG published stop short of a zero-day explosion. Exploited bugs in the first eight months of 2026 totaled 141, with 127 logged across all of 2025. The monthly exploitation average climbed from 10.5 to roughly 18, and high-risk exploitation more than doubled, from 28 last year to 75 so far, therecord.media and securityweek.com both reported.
Zero-day frequency edged upward only modestly, averaging 8 per month in 2025 and 11 in 2026, with August reaching 22. Zero-days made up 62% of observed exploitation in the period. The remainder came from n-day bugs, where adversaries pivot quickly on already-disclosed flaws.
Where AI actually shows up
GTIG believes the real shift is how fast n-days get weaponized after disclosure. One case the team highlighted involved a BeyondTrust command injection flaw, CVE-2026-1731, flagged earlier this year by federal cyber defenders and uncovered by a third-party AI research agent. Attackers were already chaining it in the wild inside days of disclosure, GTIG said.
GTIG observed a threat cluster attacking the bug inside a four-day window from public disclosure, and five more clusters showed up by day seven. Post-exploitation activity included privilege escalation, data theft, and follow-on payloads such as SNOWLIGHT, SPARKRAT, and cryptominers.
Kelli Vanderlee, senior analyst at GTIG, said the team expects AI-assisted discovery and exploitation to keep growing in the short and medium term. Attackers are using large language models to automate diff review, parse vendor advisories, and stitch together proof-of-concept code, turning patch diffs into working attacks faster than defenders can issue guidance.
Who is actually feeding the flood
A handful of vendors are responsible for a large share of the disclosures, securityweek.com noted. Totolink router firmware and Oracle both appear prominently in the totals. Many of the reported flaws live on the network edge, in appliances, remote-management tools, and exposed services, exactly where break-in is easiest.
Disclosures in AI systems themselves are also climbing fast. GTIG logged more than two thousand AI-related CVEs across January 2025 through August 2026, and the pace is still accelerating, according to the report.
What it means for you
Raw CVE volume is misleading. GTIG noted that automated CVE assignment in open-source ecosystems inflates the count — Linux kernel descriptions alone generated about 5,000 CVEs through August with no observed in-the-wild zero-day exploitation. Only about 0.23% of disclosed vulnerabilities, roughly one in 431, were seen being exploited. Half of AI-discovered bugs lead to remote code execution, versus 26% of non-AI bugs, which is a reason to weight them higher in triage.
That ratio cuts both ways. The volume is mostly noise, but the signal is sharper than it has ever been: when a high-risk CVE drops, a four-day weaponization window is now realistic. Patch SLAs need to reflect that. Say you run a perimeter stack that includes BeyondTrust or a similar remote-access tool, and your patch cadence is "next sprint." A four-day gap is now table stakes for opportunistic access brokers. Patch and validate the same day, and watch your edge telemetry for the SNOWLIGHT and SPARKRAT indicators GTIG called out.
Bottom line
Stop optimizing for CVE count. Watch exploitation telemetry, KEV listings, and CVEs in your specific stack, and treat the four-day weaponization window as the planning baseline. AI didn't invent vulnerability exploitation, but it just compressed the clock.
Key takeaways
- Monthly disclosures hit 10,740 in August; the year's exploited-bug count has already passed all of 2025.
- Exploitation is shifting toward faster n-day weaponization, with a four-day window confirmed in the BeyondTrust case.
- Only about 0.23% of disclosed bugs get exploited, so triage by exposure and stack relevance, not raw CVE volume.
- Prioritize same-day remediation for high-risk edge appliance and remote-access CVEs.
Frequently asked questions
How many vulnerabilities were disclosed in August?
Google Threat Intelligence Group recorded 10,740 disclosures in August, up from 5,045 in January.
How fast are attackers exploiting n-day bugs?
GTIG saw attackers actively chaining CVE-2026-1731 within four days of public disclosure, with five additional clusters appearing within seven days.
Are zero-days increasing?
Only marginally. Monthly zero-day exploitation averaged 8 in 2025 and 11 in 2026, with August spiking to 22. The bigger shift is faster n-day weaponization.
Should I treat all disclosed CVEs as urgent?
No. GTIG found only about 0.23% of disclosed vulnerabilities were observed being exploited. Focus on KEV-listed bugs and high-risk CVEs in your specific stack.
How is AI changing the picture for defenders?
AI is compressing the gap between disclosure and exploitation, and AI-discovered bugs more often lead to remote code execution, raising the stakes for fast triage.