Between September 24 and October 8, 2026, five announcements described the same shift. Google said an internal agent had confirmed more than 500 cross site scripting bugs in Google's own web applications. IBM and Red Hat said they had fixed over 400 flaws in common Java libraries that nobody had reported before. Anthropic opened a free AI scanner to open source maintainers and warned that an open weight model from China could build exploits end to end. And Google's threat intelligence group reported that a vulnerability an autonomous research agent had discovered was under attack four days after it became public.
The number that matters in all of this is the gap between disclosure and exploitation, because it sets how fast a patch has to move. Google's group counted 141 vulnerabilities exploited in the wild from January through August 2026, already past the 127 it recorded for the whole of 2025. Monthly disclosures doubled over the same stretch, with January's 5,045 becoming August's 10,740, and the pace of exploitation rose from roughly 10.5 flaws a month in 2025 to 18.
The hunters
Google's agent is called PageBreak. Built by the company's product security team, it ran as a pilot from November 2025 and became a full project in January 2026. Candidate bugs go to validators that fire a real payload at a running system and watch whether it lands, which Google says keeps false positives close to zero; anything unconfirmed stays inside the security team. The result was over 500 confirmed cross site scripting flaws. Applications built on Google's high assurance web frameworks fared very differently, with only 2 such bugs found across hundreds of them as of September 4.
IBM and Red Hat announced on October 6 that their Lightwell program had found and fixed over 400 bugs in common Java libraries that had never been reported, and had backported the fixes into the older versions companies actually run. Their announcement named no libraries, CVE numbers or severity ratings. Gunnar Hellekson, who runs Lightwell at Red Hat, said AI agents are indifferent to a library's age or reputation for stability, since one small crack is enough to chain an attack.
Anthropic's OSS Scanner, launched on October 8, puts a number on the bottleneck. The company said that over six months its models had turned up 29,000 candidate vulnerabilities in important open source projects, of which its staff had managed to review about 6,000. The new service sends model generated reports straight to maintainers who opt in, with no human review, each with a reproducer and often a candidate patch. In a check of 97 high or critical findings from 48 projects, penetration testers judged 85 solid enough for coordinated disclosure, 11 real but duplicated, and one invalid.
The hunted
The case Google's analysts single out is CVE-2026-1731, an unauthenticated command injection in BeyondTrust Remote Support and Privileged Remote Access that was found autonomously by Hacktron AI's research agent. Four days after it became public, one group was exploiting it; five more followed within a week, stealing data and dropping malware and cryptominers. The flaws AI finds also skew serious: remote code execution is the outcome for half of those Google could attribute to AI, against 26 percent of flaws found by other means. Google's reading is that most of the growth comes from attackers turning fresh patches into exploits, possibly with the same AI tools, since zero day exploitation only edged up from 8 to 11 a month.
The clearest sign that the attackers' side has automated too came from the Netherlands. DIVD, the Dutch vulnerability disclosure institute, was breached on September 21 through two previously unknown flaws in the Zammad ticketing system: CVE-2026-102489, which allows remote code execution in Zammad 6.3.0 through 6.5.4, and CVE-2026-102490, which turns the service account into root. DIVD said the intruder behaved like an agent deciding each next step on its own, fast and careless, even polluting its own interception attack with password spraying. Data was taken; network segmentation limited how far it got. Its advice to Zammad users: upgrade to version 7 or go offline.
The tools are also spreading. On September 29 Anthropic published its assessment of GLM-5.3, an open weight model from Zhipu AI that builds exploits end to end, succeeding in 50 of 410 attempts on a benchmark of browser engine flaws against 56 for Anthropic's own restricted model. Simple tricks got past its safeguards in 64 to 100 percent of Anthropic's trials. In one session the smaller Flash variant, given public details of a patched Chrome flaw, CVE-2026-11645, and one other known bug, produced a working exploit chain after eight hours of unattended work and 20 minutes of human attention, at an estimated $20.40 in API costs. NIST's AI standards center had already rated it, on September 17, the most capable open weight model yet for offensive cyber work.
What changes for defenders and maintainers
The flood has a cost on the receiving end. As of October 1, 2026 Google stopped accepting product vulnerability reports to its open source bug bounty, citing a surge of automated submissions, most of them invalid, and promised an update in the first quarter of 2027. The difference between that flood and PageBreak's 500 verified findings is validation: a report with a working reproducer is worth an engineer's hour, and one without is a cost.
A fix that ships without a backport for the versions people actually run leaves the gap open. Internet facing software, remote access products above all, needs updates applied in days, since every public advisory is a map that an agent can follow. For maintainers, the practical move is to accept only reports that come with a reproducer and to plan triage capacity as deliberately as development time, because the finds will keep coming faster than people can read them.
Anyone defending exposed systems should assume the BeyondTrust schedule, four days from disclosure to exploitation, from here on.