Google found exploit_agent: a framework that diffs patches, writes the exploit and fires it with no human involved
GTIG's Q4 2026 AI Threat Tracker reports the first fully automated patch-to-exploit pipeline seen on an attacker's server, plus LLM Mixers reselling stolen model access and a Claude Code settings.json backdoor. What the evidence supports and what to change.
On November 16, Google Threat Intelligence Group published its quarterly AI Threat Tracker, titled “Agentic Threat Activity Crosses a Significant Threshold”. The headline finding: for the first time, GTIG says it has identified threat actors using LLM frameworks to automate vulnerability discovery and exploitation end to end, without a human in the loop. Its previous report had described only partially automated attack sequences.
The report is worth reading past the headline. It also documents how stolen model access is being resold, and a supply chain technique that turns a coding agent’s project settings into a backdoor installer. This piece covers what GTIG actually found, how strong the evidence is, and what follows for patching and for teams that use coding agents.
What was on the server
GTIG found an exposed staging server hosting a multi-agent framework it calls exploit_agent. According to the report, the framework:
- integrates with multiple frontier and local LLMs;
- runs continuous differential analysis of popular web application software, comparing code before and after patch announcements;
- analyzes the differences to identify the vulnerability being fixed, drafts an exploit, and executes it automatically.
The report’s example shows the framework inspecting an access-control check in version 2.19.25 of Spectra (Ultimate Addons for Gutenberg), a widely installed WordPress plugin. A configuration file on the server carried Russian-language comments describing “LLM API keys for auto-exploitation (exploit_agent pilot)” and telling the operator to enter the keys in the file, “not in the chat.” That last phrase suggests the tooling itself was built with a coding assistant.
The same server held data exfiltrated from victims and scripts for what happens after a break-in: internal reconnaissance, brute-forcing, installing webshells and removing other attackers’ webshells. Victim telemetry pointed to French websites in media, entertainment and education.
Three limits are worth keeping in view. Google does not name the operator, the models used, a victim count or dates. The word “pilot” in the operator’s own comments suggests an early deployment. And what GTIG describes is a framework configured to run autonomously found next to stolen data, which is strong evidence the pipeline was used but is not the same as an observed, attributed intrusion chain. The Register describes the operator as a suspected Russian speaker, which is as far as the evidence goes.
Why attackers automate n-days first
Patch diffing is an old technique. A security fix is a precise description of the bug it closes, and skilled researchers have been turning patches into exploits for decades. What changes with agents is cost and coverage: one pipeline can watch every release of every popular plugin and CMS and try each fix as soon as it lands.
GTIG explains why this is the easier target for an LLM than zero-day hunting. The search space is finite. An agent can focus on two versions of the code, the release notes, any published research and any proof-of-concept code. Finding an unknown flaw means casting a far wider net.
The report’s numbers support that reading. Between January and August 2026, overall vulnerability disclosures and exploits nearly doubled. Zero-day exploitation rose much less, from an average of 8 a month in 2025 to 11 a month in 2026, before a spike to 22 in August.
Two cases in the report show the timeline compressing:
- PaperCut. After two print-server flaws (CVE-2026-81578, an authentication bypass, and CVE-2026-82078, unsafe dynamic class loading) were disclosed in late August, researchers and Huntress generated working exploits with LLMs within a day. GreyNoise reported a campaign four days after disclosure in which an actor likely used LLMs to build and test exploits, assemble a target list and hit hundreds of organizations, going from initial foothold to domain administrator within minutes.
- PHP. For the two PHP vulnerabilities patched on October 21 (CVE-2026-109618 and CVE-2026-109619), a proof-of-concept video and early patched code were public before the official release, so both researchers and attackers could diff ahead of the patch. Afterwards, hundreds of claimed but non-functional proofs of concept appeared, which defenders then had to verify or debunk.
That second effect is easy to miss. Cheap exploit generation raises the volume of fake exploits as well as real ones, and triage time is a cost either way.
Frameworks seen in real intrusions
Beyond exploit_agent, GTIG says multiple actors are now using open-source agentic pentesting frameworks during live intrusions. It separates what it has seen deployed from what it has seen tested:
- CAIRN, a multi-agent framework built around an observe-orient-decide-act loop with a shared blackboard, was used in intrusion activity against targets including the US aerospace sector.
- PentAGI was the base for a customized orchestration platform that one actor used between late August and early October for systematic reconnaissance and vulnerability scanning across organizations in Asia.
- CyberStrikeAI and HexStrike AI were seen in experimentation. The first connects LLMs to more than 100 security tools through a Go backend; the second exposes more than 150 tools over the Model Context Protocol.
- ARTEX is treated as already well known. Cloudflare’s network-level measurement of ARTEX forks last week covers that side in more detail.
GTIG also reports evidence that a group linked to China was developing its own automated attack framework, likely for espionage.
The report is candid about the weakness of all of these: they trade stealth for speed. Unrestricted autonomous execution generates noisy traffic, relies on recognizable identifiers and leaves code-level traces. GTIG expects actors who depend on long-term covert access to use them selectively, and criminals who do not care about being noticed to adopt them broadly, since wider reach makes up for a lower success rate. Models also hallucinate details, including targets that do not exist.
Stolen model access and a coding-agent backdoor
Two other sections deserve attention from anyone running AI workloads.
LLMjacking and LLM Mixers. Mandiant responded to several intrusions in 2026 where attackers broke into cloud environments to consume the victim’s model quota or GPU capacity. In some cases a single day of fraudulent charges passed $100,000, and the worst cases reached $1 million before remediation. Observed techniques include harvesting credentials from public GitHub repositories, creating and deleting short-lived service accounts to dodge rate limits, and exploiting CVE-2026-40217, a LiteLLM sandbox escape that turns a limited virtual key into full administrative credentials. In one case a botnet using leaked credentials generated millions of API calls in about 24 hours.
The access is then resold through what GTIG calls LLM Mixers: storefronts that route a buyer’s requests through a pool of stolen credentials and spread the volume across victims to avoid detection. Advertised prices on Russian- and Chinese-language markets ranged from $0.10 to $25 per million tokens. This matters for the first half of the report too, because it is how an operator can run frontier models inside an attack framework without an account that can be banned.
A settings file as a delivery mechanism. In intrusions against cryptocurrency firms in North America and Europe, a suspected North Korean cluster GTIG tracks as MIDNIGHT NEPTUNE gained access through fake recruiter profiles and “ClickFix” lures, then moved into internal code repositories and modified .claude/settings.json files so that starting Claude Code in the repository would trigger a download of a backdoor called SOMBERMEME. The actor also created GitHub personal access tokens named after Microsoft Copilot so they would blend in.
This is not a vulnerability in Claude Code. Project-level settings can define hooks that run commands, which is a documented feature, and a repository that many engineers open with an agent is therefore a place to plant code that runs on each of their machines. The same applies to any agent or editor that executes configuration checked into a repository.
What to change
Google sells a product in this space (AI Threat Defense is promoted in the same post), so the framing of urgency should be read with that in mind. The specific findings still point to concrete adjustments:
- Treat the patch release as the start of exploitation, not the start of a grace period. For internet-facing software with a large install base, especially CMS plugins, assume a working exploit exists within a day or two of the fix. Automatic updates for plugins and a virtual-patching layer for the gap are worth more than a monthly cycle.
- Watch pre-release channels. The PHP case shows that preview builds and public commits leak the fix before the advisory. If you ship software, that is your disclosure timeline. If you run it, it is your warning.
- Review agent configuration like code. Require review for changes to
.claude/settings.jsonand equivalent files, and alert on hooks or startup commands that fetch from the network. - Scope and meter model credentials. Per-key budgets, alerts on spend and on new service accounts, and no long-lived provider keys in repositories or CI variables. A patched LiteLLM matters if you run a gateway.
- Use the noise. GTIG says these frameworks are currently loud. Rate-based and behavioral detection on authentication endpoints and REST routes catches a lot of machine-speed probing, for now.
The open question is how long “for now” lasts. GTIG’s own expectation is that human involvement will keep falling while stealth improves as the underlying models do. The Q1 report will show whether exploit_agent was an early pilot or the first of many.
Get Insanely Good at AI
The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.
Keep Reading
How to Deploy Claude Code Auto Mode in Production
Learn how to configure Claude Code's auto mode to run unattended agent workflows, set up defense-in-depth tool guards, and manage the safety classifier.
Cloudflare: 50+ AI pentest frameworks scanned its customers in three days, and traffic is up 30x since October 1
Cloudflare's November 13 report says taking ARTEX offline produced a hydra effect: hundreds of forks, 50+ framework families and a 30x rise in agentic attack traffic. Here is what the data shows, what it leaves out, and what a stolen CI/CD token looked like in one incident.
ARTEX in the wild: Chainstack's breach and the AntiFox fork show what agentic attack frameworks actually do
Chainstack says an attacker using the ARTEX agent framework went from a default-password Grafana to production secrets in five days, and Arctic Wolf ties an ARTEX fork called AntiFox to AWS secret theft. Here is how the intrusions worked and what defenders should look for.
Operation Snowball: leaked transcripts show a jailbroken DeepSeek agent attacking a European telecom for two weeks
Malbear Labs found an attacker's open workspace with about 70 session transcripts of an autonomous intrusion harness built on DeepSeek Flash. The operator typed a few words at a time while the agent picked exploits and rewrote its tools when defenses blocked them.
One Operator, Three Agent Frameworks: Inside the $25-per-Breach AI Hacking Campaign
Reuters reported October 19 that a single hacker used autonomous AI agents against 1,200+ sites including Deere, SharkNinja, and Cummins, and a Cloud Security Alliance note reconstructs the full stack: three open-source agent frameworks running 27+ breaches at an average of $25.46 per company.