Google Finds exploit_agent Diffing Patches and Firing Exploits With No Operator in the Loop
Security / news
Google Finds exploit_agent Diffing Patches and Firing Exploits With No Operator in the Loop
GTIG's November 16 tracker reports the first fully automated discovery-to-exploit pipeline it has seen, found on an exposed staging server, plus 22 zero-days in August.

An attacker-run framework called exploit_agent compares web application code before and after a patch announcement, drafts an exploit for the difference and runs it, with no human in between, according to Google's Threat Intelligence Group (GTIG). The finding is in the AI Threat Tracker on the Google Cloud blog published November 16, 2026 by Aurora Blum, Kelli Vanderlee, Jess Xia and David Mainor.
The report calls it the first time GTIG has identified actors using LLM frameworks "to fully automate vulnerability discovery and exploitation without a human in the loop". The attacker needs nothing beyond a public patch to start. The defender's mitigation is the same as ever: install the fix before the diff becomes an exploit.
What exploit_agent does
GTIG found the framework on an exposed staging server. The report describes "continuous differential analysis of popular web application software", then analysis of the findings, exploit drafting and automatic execution "without human intervention".
Victim telemetry suggested the operator was focused on French media, entertainment and education websites. The report does not name the targeted products, the number of victims, or any CVE the framework exploited.
The design targets n-days, bugs that already have a fix. GTIG writes that attackers "may find it more efficient to use LLMs to rapidly weaponize patched n-day vulnerabilities, as they can focus research on a finite set of source materials." That is the same logic behind Project Zero's decision, covered in our report on its 60-day deadline, to drop its 30-day soak period.

August zero-days doubled the 2026 average
The tracker puts zero-day exploits at an average of 8 a month in 2025 and 11 a month in 2026 through July, then 22 in August. It says overall disclosures and exploits "nearly doubled between January and August 2026". That doubling refers to the overall count, not zero-days alone, which rose only slightly until the August spike.
- 2025 monthly average8 exploits
- 2026 monthly average, through July11 exploits
- August 202622 exploits
Source: GTIG AI Threat Tracker, Google Cloud Blog, 16 November 2026
The report does not attribute the August figure to AI tools, and one month is a thin basis for a trend.
The other frameworks GTIG names
| Framework | What the report says | Where seen |
|---|---|---|
| CAIRN | OODA-loop multi-agent framework | Intrusions against US aerospace targets |
| PentAGI | Used to scan organisations | Asia, late August to early October 2026 |
| CyberStrikeAI | Connects LLMs to 100+ security tools | Listed as an offensive framework |
| HexStrike AI | MCP-integrated harness with 150+ tools | Listed as an offensive framework |
| ARTEX | Widely publicised pentesting tool | Listed as an offensive framework |
ARTEX is the tool the FBI described in a flash advisory that The Terminal covered. GTIG's view is that these open-source frameworks automate discrete phases of live intrusions. It also says they "currently trade stealth for velocity" because autonomous execution leaves forensic traffic, recognisable identifiers and code-level traces.
What this does not show
GTIG found the staging server, not a confirmed breach chain, and gives no count of compromised sites. It also notes a cost for defenders: LLM-generated exploits "claimed to work but don't" are creating extra verification work.
The report cites CVEs it ties to wider activity, including CVE-2026-40217, a LiteLLM sandbox escape that raises "Virtual Keys" to admin credentials, and the chained PaperCut pair CVE-2026-81578 and CVE-2026-82078 for unauthenticated remote code execution. The report does not say exploit_agent used any of them.
MediaNama quotes GTIG's John Hultquist, Chief Analyst, saying "We will have to embrace agentic defense." The practical step for operators is shortening the gap between a vendor patch notice and deployment on internet-facing web applications. Google's wording is that the tools are moving "further along the autonomous scale", so the next tracker is the one to check.
Sources
More in Security
- 01Java4Shell Scores 9.2 From Its Finder and 8.1 From OpenJDK, and Needs an AVX-512 CPUCalif says a 65-character Base64 string overwrites 62 bytes of Java heap and reaches code execution in Keycloak, Jenkins and Jira. A second security firm says the source shows only a crash.
- 02CISA Gave 9 of 16 November KEV Entries a Three-Day Deadline With Forensic Triage, and 7 Got 14 DaysThe catalog's own due-date and forensicTriage fields show two tiers under BOD 26-04. Sorting November's additions by tier shows which exploited bugs CISA treats as already-compromised-until-proven-otherwise.
- 03Project Zero Cuts Its Disclosure Deadline to 60 Days on Feb. 2, 2027, Then 30 in 2028Google's bug-hunting team is dropping the 30-day patch soak period and counting every deadline from the day it reports, citing AI-assisted exploitation.
- 04Mandiant: LLMjacking Runs Up $100,000 a Day on Stolen Cloud Keys, and Logs Are Off by DefaultGoogle's Nov. 2 report says single-token test calls evade volume alerts, and it describes the LiteLLM bug differently from BleepingComputer.