CyberPeople

Hackers Taught AI Agents to Turn Security Patches into Exploits

Hackers Taught AI Agents to Turn Security Patches into Exploits

Google Threat Intelligence recorded the first fully automated AI agent that turns security patches into working exploits — with no human in the loop.

A first: an AI agent turns patches into exploits on its own

Google Threat Intelligence Group (GTIG) and Mandiant published their fourth-quarter 2026 AI threat tracker under the title "Agentic Threat Activity Crosses a Significant Threshold." The headline finding: attackers used an LLM framework to fully automate vulnerability discovery and exploitation for the first time — "without a human in the loop."

This is not about someone typing a prompt into a chatbot and copying the code out. It is about an autonomous system that watches for software updates on its own, compares code before and after a patch, identifies the fixed vulnerability, drafts an exploit, and executes it against targets — while the operator sleeps. The difference is fundamental: AI used to assist a human; now it runs the entire attack chain on its own.

Until this quarter, researchers had only documented partial automation — scanning or generating isolated code snippets. A fully autonomous "patch → exploit → attack" pipeline deployed against real targets had not been observed. That threshold has now been crossed.

How we got here: from prompting to autonomy

This is not a sudden leap but a logical stage in an evolution GTIG has documented for several quarters. In earlier reports (notably "From Prompting to Autonomy: The Evolution of Adversarial AI"), researchers described a gradual progression: first attackers used LLMs as a smarter search engine and text generator for phishing, then to write individual scripts and analyze code, and only later to coordinate multi-step attacks.

The key difference of the new stage is autonomy. Before, the model was a tool in the operator's hands: a human set the task, checked the output, corrected mistakes. Now the agent plans each step itself, distributes work across workers, and builds or installs whatever it lacks. As Elastic researchers note, ARTEX's agent acts in bursts of two to 25 seconds — a pace unreachable for a human working manually.

Another sign of maturity is the emergence of ready-made, documented, and widely replicated open-source frameworks. You no longer need to write one yourself: clone the repository, configure an API key, and run it.

What exploit_agent is and how it works

What exploit_agent is and how it works

The framework, which researchers dubbed exploit_agent (after the name found in its configuration file), was discovered on an exposed staging server. It integrates with multiple frontier models and locally hosted large language models at once.

The core technique is differential patch analysis. The system continuously compares the source code of popular web applications before and after a security update is released. When a patch changes a particular function, that is a marker: that is where the vulnerability was. The agent then analyzes the finding, drafts an exploit, and executes it automatically.

Merritt Baer, a former deputy CISO at AWS, captured the shift precisely: "Attackers didn't invent a new kind of bug. They automated the tedious part: reading the patch, finding what it fixed, and writing the exploit. A skilled person might spend days on that. Now an agent does it while its operator sleeps."

In other words, the most labor-intensive and highly skilled part of the job — reverse-engineering security updates — is now automated and scalable.

Who is behind it and who was targeted

Google does not attribute the framework to a specific group or country. However, the configuration file contained Russian-language comments describing exploit_agent as a "pilot" and instructing the operator to enter LLM API keys directly into the file, "not in the chat." The Register calls the operator a "suspected Russian-speaking operator."

The same staging server held data stolen from victims, along with several post-compromise scripts for internal reconnaissance, brute-forcing, installing webshells, and even removing webshells left by rival attackers.

Victim telemetry suggests the operator was specifically interested in French websites associated with the media and entertainment industries, as well as the education sector. Google also published a code excerpt showing the framework examining access controls in version 2.19.25 of the Spectra WordPress plugin — but did not tie it to a successful exploit.

More than one framework: ARTEX, CAIRN, CyberStrikeAI, PentAGI

More than one framework: ARTEX, CAIRN, CyberStrikeAI, PentAGI

exploit_agent is not the only tool. GTIG and Mandiant record that attackers are actively weaponizing open-source multi-agent penetration-testing frameworks:

  • ARTEX — an event-driven planner with parallel workers. Microsoft attributed its use to the group Storm-0940 in the ConnectWise compromise chain.
  • CAIRN — delegates work to existing coding agents running inside Kali containers.
  • CyberStrikeAI — a complete operator console that builds its own implants and runs C2 listeners and a webshell manager.
  • PentAGI — lowers the barrier to entry for novices. Between late August and early October 2026, GTIG tracked an attacker using a customized platform built on PentAGI for systematic reconnaissance and scanning of organizations in Asia.

By Elastic's estimate, five such frameworks (ARTEX, CyberStrikeAI, Cairn, PentAGI, and HexStrike) have drawn more than 60,000 GitHub stars combined. This is not exotic — it is widely available tooling that installs with a single script or Compose file.

Alongside these sit companion services: ZeroHollow (an inference service for exploit development and LLM routing) and Agent Zero (an open-source agent framework with autonomous code execution and dynamic tool generation).

Tellingly, these frameworks already leave recognizable traces in telemetry that defenders can hunt for. ARTEX, for example, runs scanners such as nuclei, httpx, katana, ffuf, and sqlmap with three parallel workers by default and sends requests with the User-Agent artex-enrich/1.0. CyberStrikeAI is known for webshells carrying the User-Agent Mozilla/5.0 (compatible; CyberStrikeAI-WebShell/1.0) and an implant that persists via a scheduled task named SystemUpdate. PentAGI connects from compromised hosts to TCP ports 28000–29999. Automation, in other words, is not only an advantage for attackers — it also makes them more predictable to those who can read network telemetry.

Why it matters: the window between patch and exploit is vanishing

Traditional defense logic relies on a "window of exposure": a vendor ships a patch, and defenders have days, sometimes weeks, before attackers reverse-engineer the update and write an exploit. That is why the standard advice for years was "patch quickly."

That window is now shrinking to hours. Google states plainly that attackers use LLM pipelines to automate "labor-intensive tasks like vulnerability analysis and proof-of-concept weaponization," which "collapses the window between patch disclosure and active exploitation."

A telling example is the PaperCut NG/MF flaws in September 2026, when an attacker used hundreds of AI agents and automated differential patch analysis to rapidly weaponize n-day vulnerabilities.

There is another important shift: GTIG believes attackers may now find it more efficient to target n-day vulnerabilities than to hunt for zero-days. After all, the source material for an already-patched flaw is finite: compare code versions, read the release notes, find the POC — done. A full zero-day hunt demands far more resources and time.

The adjacent problem: stealing compute for LLMs

Alongside the frameworks grows a shadow economy around access to LLMs. GTIG documents so-called LLMJacking — monetizing compromised cloud models. Attackers harvest session tokens and platform credentials directly from victim endpoints and CI/CD environments to run their own LLM pipelines.

Underground forums and marketplaces now sell verified frontier-model accounts, tokenized API quotas, and turnkey infostealer configurations. Some attackers go further, renting cloud instances with powerful GPUs and requesting quota increases to deploy their own LLM tooling for automating further intrusions.

How to defend

How to defend

Google's report is not a reason to panic, but it is a clear signal to reorder defense priorities:

  1. Shorten update times. If the window between patch and exploit is now measured in hours, your update speed must also be measured in hours, not weeks. Automate patch management.
  2. Watch for anomalies in telemetry. Agentic frameworks leave characteristic traces: specific User-Agents, sudden new accounts, unusual connections. Elastic published ready-made ES|QL queries to detect ARTEX, CyberStrikeAI, Cairn, PentAGI, and HexStrike.
  3. Protect credentials and CI/CD. A large share of new attacks is token and key theft. Minimize long-lived secrets, enable multi-factor authentication, and monitor anomalous cloud resource usage.
  4. Check whether your data has leaked. Regularly check for breaches — our data breach check tool helps you see whether your credentials are already circulating online.
  5. Invest in team training. The faster attackers automate, the more security hygiene matters at every level. The fundamentals are in our cybersecurity training guide.
  6. Treat n-days like zero-days. Since attackers now weaponize already-patched flaws at scale, prioritize updates for products that already have public POCs or exploitation reports. The days of "we can wait to update" are over.

Google itself responds with Google AI Threat Defense (AITD) — an architecture that combines the reasoning of Gemini and other frontier models, risk prioritization from Wiz, and automated remediation from Gemini and CodeMender.

What this means for Ukrainian organizations

The report discusses French websites, but the threat itself is global and does not distinguish between countries. Open-source frameworks like ARTEX or PentAGI are available to anyone: the same toolkit an attacker used against media and education institutions in France can be aimed at Ukrainian media, education platforms, IT companies, and government services.

Those who are slow to update, or who run popular CMS platforms and web applications without proper monitoring, are especially exposed. Popularity is precisely what makes a product an attractive target for differential patch analysis: the more installations, the wider the field for automated exploitation after every update.

The practical conclusion for Ukrainian teams stays the same but grows stricter: automate updates, centralize telemetry where it can be analyzed, and do not assume that "nobody knows about us." In the age of agentic attacks, the anonymity of small organizations is no longer a defense.

Conclusion

The threshold has been crossed: we have moved from "AI helps write phishing emails" to "AI turns patches into exploits and attacks on its own, without a human." That does not mean defending is pointless. It means the old metrics — "we can update within a week" — no longer work.

Google puts the forecast bluntly: "Given the low barrier to entry, ease of integrating new exploits, and massive scaling potential, we believe it is highly likely that threat actors will continue to use these frameworks in campaigns targeting organizations globally."

Those who manage to automate their defense as quickly as attackers automated their offense will win. Everyone else risks becoming a target that an agent attacks faster than the team can finish reading the patch report.

Stay ahead of threats

Weekly cybersecurity intelligence in your inbox. No spam.

CyberPeople contributor