← Blog
AI AgentsAutonomous ExploitationPatch DiffingThreat Intelligence

AI Agents Now Build Exploits From Patches With No Human in the Loop

Google's GTIG found exploit_agent — an autonomous framework that diffs each software patch, writes the exploit, and fires it with no human in the loop.

Zero Hunt Research··7 min read

Published by Zero Hunt, an autonomous AI red team on an on-premise appliance running private AI: automated penetration testing for networks and infrastructure, black-box or gray-box, with a human approving every step that matters.

Google's Threat Intelligence Group found an exposed staging server running a framework it calls exploit_agent. The framework watches popular web applications for security patches, compares the code before and after the fix, works out what the patch is quietly closing, writes an exploit for it, and fires that exploit at real targets — with no person approving any step. This is the thing the industry has been saying was a year or two away. GTIG published it on November 16, 2026, and its framing is blunt: adversaries have, for the first time, been seen fully automating vulnerability discovery and exploitation without a human in the loop.

Developing story — first published 18:29 CET (17:29 UTC), November 16, 2026. Updated as Google, Mandiant and the affected vendors publish more.

The patch is the weapon: inside exploit_agent

A security patch is a disclosure. The diff between the vulnerable release and the fixed one tells anyone who reads it exactly where the bug was and roughly how to reach it. Skilled attackers have reverse-engineered patches by hand for decades — that is what an "n-day" exploit is. What changed is who is doing the reading.

GTIG describes exploit_agent as "a multi-agent vulnerability research and weaponization framework" that integrates multiple frontier and local LLMs. Its job, in Google's words, is to:

conduct continuous differential analysis of popular web application software. It compares code before and after patch announcements, analyzes findings to identify potential vulnerabilities, drafts exploits, and automatically executes them without human intervention.

On the same server GTIG found the agent's output: data exfiltrated from victims, plus scripts for internal reconnaissance, brute-forcing, installing web shells, and — a detail that tells you how crowded this space has become — removing other attackers' web shells. An environment file carried the comment "LLM API keys for auto-exploitation (exploit_agent pilot)." One captured run showed the framework analyzing an access-control flaw in the Spectra (Ultimate Addons for Gutenberg) WordPress plugin, version 2.19.25. Victim telemetry pointed at French media, entertainment and education sites. Google did not attribute the framework to a named group or country.

The point is not that exploit_agent is sophisticated. It is that it is tireless. A human patch-differ sleeps; this one runs continuously against every popular release it can watch.

AI agents that build exploits from patches, ranked by what Google actually saw

exploit_agent is one custom build. The broader shift GTIG reports is that off-the-shelf agentic offensive frameworks are now in real intrusions, not just demos. The report is careful about the difference between observed in the wild and seen in testing — a distinction most coverage flattens, and one that matters when you decide what to defend against first.

Framework What it is Google's observation
CAIRN Multi-agent pentest framework; OODA loop, agents share a "blackboard" Observed in the wild, incl. against US aerospace
PentAGI Autonomous offensive framework; lowers the barrier for entry-level operators Observed in the wild; recon + scanning against orgs in Asia, late Aug–early Oct 2026
ARTEX Go-based multi-agent LLM pentest + C2 platform, six sub-agents Actively leveraged; tracked as UNC7228, overlapping Microsoft's PRC-nexus STORM-0940
CyberStrikeAI LLM-native "single pane of glass" C2 wiring an LLM to 100+ tools Experimentation only
HexStrike AI MCP harness exposing 150+ offensive tools to 12+ sub-agents Experimentation only

ARTEX is the one that has already drawn separate GTIG and Mandiant reporting: its original GitHub repository was pulled by its own author in October 2026 because of how widely threat actors had picked it up, and it now spreads through forks and a paid "ARTEX PRO" build. We covered an earlier ARTEX campaign in detail here. The through-line across all of them is the one exploit_agent makes literal: the slow, expert parts of an intrusion — reading a diff, writing a working proof-of-concept, deciding where to pivot — are being handed to agents.

Why patch diffing at machine speed collapses the window

Zero-day counts barely moved: GTIG notes monthly zero-day exploits rose from a 2025 average of 8 to 11 in 2026, spiking to 22 in August. The dangerous curve is the n-day one — the gap between a patch shipping and an exploit being usable. That gap is what AI compresses.

Two cases from the report show the mechanism without any agent at all:

  • PaperCut. In late August, CVE-2026-81578 (an unauthenticated access-control bypass) and CVE-2026-82078 (unsafe dynamic class loading) were chained into unauthenticated remote code execution. GreyNoise saw a campaign against hundreds of organizations four days after disclosure, with attackers going from foothold to domain administrator within minutes.
  • PHP. When CVE-2026-109618 and CVE-2026-109619 were patched on October 21, a video proof-of-concept and early patched code appeared, letting researchers and attackers compare versions — and hundreds of (mostly non-functional) PoCs followed within days.

Now put an agent on that loop. GTIG states plainly that adversaries are automating "differential patch examination, dynamic protocol fuzzing, and proof-of-concept (POC) weaponization," including by "comparing source code versions (including from pre-release or preview channels)." The attacker's AI gateway itself is a target: CVE-2026-40217 is a LiteLLM sandbox escape that promotes a scoped key to admin. The patch window that your risk model assumes is weeks long is, against a framework that reads the diff the hour it drops, a few hours at most.

They trade stealth for velocity — and that is your opening

Here is the part worth holding onto. GTIG is explicit that these tools "currently trade stealth for velocity." Unrestricted autonomous execution "generates forensic traffic, relies on recognizable identifiers, and leaves behind distinct code-level traces." ARTEX runs come in bursts of roughly 2 to 25 seconds and leave telltale debris — web shells named like cl1217.jsp, files prefixed artex-probe-, accounts containing ArtexProbe.

That noise is a gift, but only if you are watching the right surface. Google's own guidance is that "tool names, hashes, IP addresses, and other static indicators … will not cover much of this activity." Signatures chase yesterday's agent. The durable signal is behavioral and volumetric: an unprecedented rate of control-plane actions, a host that suddenly fans out to internal targets, bursts of egress to webhook and tunneling services.

What to do now

No new patch is owed to you here — the fix is operational, not a KB number:

  1. Shorten your own clock to match theirs. Treat "patch available" as "exploit imminent" for internet-facing web apps, plugins and gateways. Prioritize by real exposure, not CVSS alone — the BOD 26-04 risk-based tiers are a sane model even if you are not a US federal agency.
  2. Inventory what reads a diff for you. AI gateways (LiteLLM and peers), app-builder platforms and MCP endpoints are now initial-access targets. Confirm authentication, kill default keys, and block server-side request forgery to cloud metadata.
  3. Hunt behavior, not hashes. Baseline control-plane velocity and server egress. Alert on machine-tempo bursts of reconnaissance, on hosts reaching webhook/OAST/tunneling infrastructure, and on fan-out a host has no business performing.
  4. Validate before you assume you are safe. A version string below the fixed release tells you that you should patch. It does not tell you whether a given n-day is actually drivable to code execution on your build and configuration. Those are different questions.

What is not known yet

GTIG did not attribute exploit_agent to a named actor or country, and its account is drawn from one exposed staging server — representative of the operator's intent, not a measured count of victims. Whether these frameworks reliably produce working exploits across targets, rather than many failed attempts, is still unclear; the PHP wave of non-functional PoCs is a reminder that "an agent wrote an exploit" and "the exploit works" are not the same claim. Expect the picture to sharpen as Mandiant and the affected vendors publish.

The symmetric answer runs on your side of the wire

If an autonomous framework can read a patch and write the exploit before you have finished deploying it, the only honest way to know your exposure is to run the same process against your own estate first — governed, on-premise, and on your schedule. That is what Zero Hunt is: an autonomous AI red team that runs on an on-premise appliance on private AI (our own models, ZeroHunt Apex and Apex Pro), with a human in the loop for every action that matters. Its Generative Zero-Day PoC and Exploit Forge do, under your control, what exploit_agent does against you — identify the exact installed version, analyze that version's source for input-reachable weaknesses (gray-box, against your own gear), and let a sealed coding agent iterate a runnable proof-of-concept in an ephemeral hardened container that never touches the host, until it demonstrably works or demonstrably cannot. Because the models sit in your rack, your source and credentials never travel to an external model — the kind of external model that, elsewhere in this same report, is itself the thing diffing patches to attack you. A change-triggered campaign can fire the moment a new asset or patch appears, which is roughly the cadence the adversary now runs at.

When the n-day window is measured in hours, you cannot patch everything at once, and that is where the AI Remediation Advisor earns its place: it ranks fixes by real exploitability — CISA KEV first, then what your own campaign actually proved reachable, then CVSS and EPSS — hands you the exact fix with rollout and rollback notes and a compensating control for the hosts you cannot touch tonight, then re-verifies with the same proof that demonstrated the problem. The attackers have automated the race. The defensible move is to automate your side of it, continuously and on your own terms — not to keep patch-diffing by hand against something that never sleeps.

Is this exploitable in your environment?

Zero Hunt answers that on your own network: an autonomous AI red team on an on-premise appliance, running on private AI, black-box or gray-box, with a human approving every step that matters. Proof of what is exploitable, the fix, and signed evidence — no data leaves your perimeter.