Google finds exploit_agent, an AI framework that exploited flaws with no human in the loop
Google Threat Intelligence Group said on 16 November 2026 it found exploit_agent, an LLM tool that diffs patches, writes exploits and runs them with no human.
By Berkant A.Published
Tech AI LatestGoogle Threat Intelligence Group and Mandiant reported on 16 November 2026 that they found, for the first time, attackers using an LLM framework to automate vulnerability discovery and exploitation with no human in the loop. The tool, dubbed exploit_agent, sat on an exposed staging server, compared web application code before and after patches, drafted exploits and ran them. Victim data pointed to French media, entertainment and education websites.
Key facts
- Published: 16 November 2026 in the GTIG AI Threat Tracker (Google Cloud blog)
- Finding: exploit_agent, a multi-agent framework using frontier and local LLMs, ran patch diffing, exploit drafting and execution without human intervention
- Targets: telemetry on the server pointed to French media, entertainment and education websites
- Other frameworks seen in the wild: CAIRN and PentAGI; experimentation with CyberStrikeAI and HexStrike AI
- LLMJacking: single-day fraudulent charges above $100,000 in some Mandiant cases, up to $1 million cumulative
- Not stated: who operates exploit_agent or how many sites it compromised
What Google found
Google Threat Intelligence Group (GTIG) and Mandiant published the November edition of their AI Threat Tracker on 16 November 2026, titled "Agentic Threat Activity Crosses a Significant Threshold". The central claim is a first: in the past quarter, GTIG says, it identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop.
The evidence is a framework GTIG calls exploit_agent, found on an exposed staging server. Google describes it as a multi-agent system that connects to several frontier and local models and runs continuous differential analysis of popular web application software. It compares code before and after patch announcements, works out which change fixed a vulnerability, drafts an exploit and executes it automatically. The server also held data taken from victims and scripts for internal reconnaissance, brute forcing, installing webshells and removing other attackers' webshells.
Google published one example of the framework's working files: an analysis of an access control check in version 2.19.25 of the Spectra WordPress plugin. A configuration file carried Russian-language comments labelling the API keys as being for the "exploit_agent pilot". Victim telemetry suggested the operator was interested in French websites in media, entertainment and education. Google does not name the operator or say how many sites were hit.
The wider agentic toolset
exploit_agent is one part of a broader shift the report describes. GTIG says it has independently observed the open source frameworks CAIRN and PentAGI deployed in real intrusions, and seen experimentation with CyberStrikeAI and HexStrike AI, alongside ARTEX, the framework Microsoft linked to Storm-0940 in October. CAIRN was used against targets including the US aerospace sector. One actor used a customised PentAGI build for reconnaissance and scanning across Asia between late August and early October 2026, and GTIG says a group with a PRC nexus was developing its own attack framework.
The report ties this to the shrinking window between a patch and an attack. It cites the PaperCut flaws CVE-2026-81578 and CVE-2026-82078, for which researchers and Huntress produced working exploits with LLMs within a day of disclosure, and a GreyNoise-reported campaign four days after disclosure. Google adds a cost on the defensive side: after the PHP flaws CVE-2026-109618 and CVE-2026-109619 were patched on 21 October, hundreds of non-functional proof-of-concept exploits, likely LLM-generated, were published, and each had to be checked.
GTIG also says the frameworks currently "trade stealth for velocity": unrestricted autonomous execution leaves forensic traffic, recognisable identifiers and code-level traces. That makes them easier to detect today, and Google expects stealth to improve as models do.
Stolen AI access and coding agents as a way in
The report gives new figures on LLMJacking, where attackers use stolen cloud credentials to run AI models at the victim's expense. Mandiant says some 2026 incidents produced single-day fraudulent charges above $100,000 and, in the worst cases, cumulative abuse of $1 million before remediation. One actor exploited CVE-2026-40217, a sandbox escape in LiteLLM that turns limited inference keys into admin credentials. Underground "LLM Mixer" services resell this access at between $0.10 and $25 per million tokens, according to GTIG. The trend matches the nine AI platform breaches Wiz counted in October.
Coding agents are now also an intrusion path. GTIG says a suspected North Korean cluster it tracks as MIDNIGHT NEPTUNE, after phishing cryptocurrency staff with fake recruiter profiles, edited .claude/settings.json files in internal repositories so that starting Claude Code would download a backdoor called SOMBERMEME.
What changed and what to watch
In its September tracker, GTIG described attackers moving from prompting toward agent campaigns, and its vulnerability trends report said disclosures had doubled in 2026. The November report moves from partly automated pipelines to an observed one with no human step. For defenders the practical reading is that the time to patch internet-facing web software after a fix is published keeps shrinking, which is also why Project Zero cut its disclosure deadline.
Google says SecOps has rulepacks to detect ARTEX and that Mandiant released a hardening guide for LLM-enabled attacks. Open questions are who runs exploit_agent, which models it called and whether model providers have cut off its keys. More AI security coverage is on the cybersecurity hub.
What changed
| Before | GTIG had reported partly automated, LLM-assisted vulnerability scanning and exploitation, with humans still directing the steps |
|---|---|
| After | GTIG reports the first observed framework, exploit_agent, that finds, weaponises and runs exploits from patches without human intervention |
| Who it affects | Operators of internet-facing web applications and WordPress sites, security teams on patch timelines, companies with cloud AI credentials |
Questions people ask
- What is exploit_agent?
- It is the name Google Threat Intelligence Group gave to a multi-agent framework found on an exposed staging server. It compares web application code before and after patches, drafts exploits with frontier and local LLMs and runs them without a human, according to Google's 16 November 2026 report.
- Who was targeted?
- Victim telemetry on the server suggested French websites in media, entertainment and education. Google did not name the operator or give a count of compromised sites.
- What should defenders do?
- Google's report stresses that the gap between a patch and exploitation is shrinking, so patching internet-facing web software quickly matters more. Google points to its SecOps detection rules for ARTEX and a Mandiant hardening guide.
Sources
- GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold (Google Cloud, 16 Nov 2026)
- Google GTIG warns AI agents automate exploits, steal credentials and scale cybercrime (Security Affairs, 17 Nov 2026)
Reference pages
- Google company
Spotted an error? See the corrections policy and tell the editor. Facts in this story come from the sources listed above; vendor claims are attributed.
Comments
Loading comments.




