How fast can AI turn a patch into a working exploit?
GTIG's November 16 report names exploit_agent as the first fully automated case it has seen. The PaperCut campaign it cites began four days after the advisory.

Google Threat Intelligence Group reported on November 16 that it found exploit_agent, a framework that diffs patched code and runs the exploit it writes with no human involved. In the PaperCut campaign the report cites, GreyNoise measured just under four hours from an empty workspace to code execution on a real victim.
What did Google actually find on that server?
GTIG found a staging server its operator had left exposed, running a framework it calls exploit_agent. The report, published on November 16 by Aurora Blum, Kelli Vanderlee, Jess Xia and David Mainor, describes a multi-agent setup wired into several frontier and locally hosted models. It watches popular web application software, compares the code before and after each patch announcement, writes an exploit for what it finds, then runs it unattended.
Data taken from victims sat on the same server, with scripts for reconnaissance, brute forcing and installing webshells. Victim telemetry on the server suggested the operator was going after French sites in media, entertainment and education. GTIG says this is the first time it has seen discovery and exploitation fully automated, in its phrase, "without a human in the loop."
Which products did these automated pipelines hit?
Three are named, and all three are in ordinary enterprise use. The code exploit_agent was caught reading is the access control logic in Spectra, the WordPress plugin published as Ultimate Addons for Gutenberg, at version 2.19.25, released on April 27, 2026. The current release is 2.20.6, dated November 3, and the plugin runs on more than a million sites. Seven releases since 2.19.25 carried a security fix, credited to Wordfence, WPScan, Patchstack and two named researchers.
The second is PaperCut NG/MF, through CVE-2026-81578, an authentication bypass, and CVE-2026-82078, an unsafe dynamic class loading flaw. Chained, Huntress reports, they give remote code execution before any login. The third is PHP, where CVE-2026-109618 and CVE-2026-109619 were patched on October 21.
How long did the PaperCut attack really take?
GTIG puts four days between the advisory and a campaign against hundreds of organizations. PaperCut published its advisory on August 27. Researchers and Huntress had working exploits inside a day, Huntress reproducing the chain against an unmodified PaperCut NG 25.0.11.75758 server. GreyNoise logs the operator starting work on August 31 and the agents launching the campaign at 08:30 UTC the next morning.
GreyNoise's own measurements are the ones to take to a change board. From an empty workspace to remote code execution on a real victim took just under four hours, with the first domain administrator two hours later. At least eleven organizations fell inside 26 seconds. GreyNoise counted at least 440 instances at 395 identified victim organizations in 48 countries, and says others could not be attributed. Domain administrator was reached at 12 of them, in five to 144 minutes from foothold; elsewhere it saw multiple-day gaps. CISA added both CVEs to its catalog on August 31 with a due date of September 14, and GreyNoise's last logged code execution was September 2, so the campaign was over inside the first days of that window.
Is the exploit_agent case the same as the PaperCut campaign?
No, and the difference matters when you read the coverage. GreyNoise's timeline records the PaperCut operator being "prompted for permission to continue," so a person was in that loop. GTIG's finding is narrower and newer: exploit_agent is the first case it has seen with nobody in the loop at all, and GTIG names no Spectra victim and no CVE for it. What the two share is the method, differential analysis of a vendor's own patch, and the limits: in at least one instance Cloudflare's web application firewall defeated the PaperCut attacker, and GreyNoise's conclusion was that "hardening of environments still matters" against attackers using AI.
What should IT teams do about AI-built n-day exploits today?
- Patch PaperCut to 26.0.5, 25.0.13 or 24.1.10. Huntress found 47 percent of the roughly 2,500 installations it tracks sitting on v23 or older, builds for which, it says, "no patch is currently available." That is about 1,175 servers with nothing to install.
- Take the PaperCut application server off the public internet whatever version it runs. Huntress recommends restricting web access to trusted addresses or putting it behind a VPN. GreyNoise notes these servers run as SYSTEM on Windows by default and are usually domain-joined.
- Check your WordPress estate for Spectra below 2.20.6 and update. Seven security releases have shipped since the version GTIG saw.
- Treat the CISA catalog due date as the deadline, not a day count. Entries now point to BOD 26-04 and carry a per-CVE date, and for these two it was 14 days after they were added.
- Put a web application firewall in front of anything internet-facing you cannot patch this week. It is the one perimeter control GreyNoise recorded defeating this attacker outright.
Sources
- Google Threat Intelligence Group: GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold, 16 November 2026
- GreyNoise: Agents Gone Wild: An AI-Orchestrated Global Campaign Against PaperCut NG/MF, 9 September 2026
- Huntress: PaperCut actively exploited, 27 August 2026
- CISA: Known Exploited Vulnerabilities catalog, version 2026.11.13
- WordPress.org: Spectra Legacy (Ultimate Addons for Gutenberg) plugin page and changelog
- SiliconANGLE: Google says an AI framework turned security patches into exploits with no human in the loop, 16 November 2026