Skip to content
audiotale.net
Sign in
NewsNews

Google GTIG: exploit_agent turns patches into exploits, no human involved

Google's threat team says attackers ran an AI framework that diffs patches, writes exploits and runs them automatically, the first fully autonomous pipeline it has seen in the wild.

audiotale.net team

· 8 min read

Written by AI from the sources linked below. How we use AI

On this page
  1. Key points
  2. What to do
  3. What happened
  4. Background
  5. The numbers
  6. How the frameworks work
  7. What it means
  8. What we don't know yet
  9. What to watch next
  10. Sources

Google's Threat Intelligence Group (GTIG) says it has found the first case of attackers running an AI framework that turns security patches into working exploits with no human in the loop. The tool, which the operator called exploit_agent, sat on an exposed staging server next to data stolen from victims. The finding, published on 16 November 2026 in GTIG's fourth-quarter AI threat tracker, matters because it shortens the time defenders have between a patch appearing and an exploit existing for it.

Key points

  • GTIG says that for the first time it has seen threat actors fully automate vulnerability discovery and exploitation with no human involved, using a multi-agent framework named exploit_agent.
  • The framework compares a web application's code before and after patch announcements, drafts exploits for the differences and runs them automatically, according to GTIG.
  • Victim telemetry on the server points to French media, entertainment and education websites; GTIG describes the operator as using Russian-language tooling.
  • GTIG's own earlier report, two months ago, said it had not yet seen fully autonomous attack pipelines used against real targets.
  • Open source agent frameworks (ARTEX, CAIRN, CyberStrikeAI, HexStrike AI, PentAGI) are already being used in live intrusions, GTIG says, though they are noisy and easier to detect.

What to do

This is a research report, not a single product advisory, so there is no one version to install. The practical lessons from GTIG's findings:

  • Shorten your patch window for internet-facing software. The report's core point is that an attacker can now start from the patch itself. Prioritise fixes for web applications, plugins and servers that face the internet, and do not wait for a public proof of concept to appear.
  • Check what runs in your repositories. GTIG describes a suspected North Korean cluster, which SiliconANGLE says is tracked as UNC6836, editing .claude/settings.json files in compromised internal code repositories so that starting Claude Code downloaded a backdoor. Treat agent configuration files in repositories like build scripts: review changes to them and restrict who can write them. This is our reading of the report, not a Google recommendation.
  • Look for the noise. GTIG says unrestricted autonomous execution leaves forensic traces. Google Security Operations has rule packs for ARTEX, and Mandiant has published a hardening guide for defending against LLM-enabled offensive tooling (both linked from GTIG's report).
  • Guard your LLM credentials. The same report covers "LLMJacking", where stolen cloud and API credentials are used to run models at the victim's expense. Rotate keys found in public repositories and alert on sudden spikes in model usage.

What happened

GTIG's report is titled "Agentic Threat Activity Crosses a Significant Threshold". Its central finding is that in the past quarter the group identified an exposed staging server that hosted a multi-agent vulnerability research and weaponization framework, which it dubbed exploit_agent. According to GTIG, the framework integrates with several frontier and local language models and is set up to carry out continuous differential analysis of popular web application software. It compares code before and after patch announcements, analyses the differences to identify potential vulnerabilities, drafts exploits and executes them automatically without human intervention.

The server also held data taken from victims and several scripts for later stages of an intrusion: internal reconnaissance, brute-forcing, installing webshells and removing other attackers' webshells. GTIG says victim telemetry suggests the operator was specifically interested in French websites associated with the media and entertainment industries and with education. The report shows a code excerpt the agent was inspecting: an access-control check in version 2.19.25 of the Spectra (Ultimate Addons for Gutenberg) WordPress plugin. A comment in the framework's environment file, in Russian, tells whoever sets it up to enter the LLM API keys themselves, "right here, not in the chat". The Register calls the operator a suspected Russian-speaking actor; GTIG's own wording is about the Russian-language tooling and comments, and it does not name a group.

GTIG does not say which vulnerabilities exploit_agent found or how many victims it compromised, and the report does not give a count of victims.

Background

Two months ago GTIG's previous tracker described attackers using AI for partly automated scanning and exploitation, but said it had not observed fully autonomous attack pipelines deployed against targets in the wild, as SiliconANGLE notes in its summary. The new report moves that line.

GTIG argues that attackers may find it more efficient to point models at n-day vulnerabilities (flaws that already have a patch) than to hunt for zero-days. The source material for a patched flaw is finite: the two code versions, release notes, security research and any proof of concept. Searching for a zero-day means casting a much wider net.

The report gives recent examples of how fast that can go:

  • PaperCut. In late August 2026 attackers exploited two PaperCut print management server flaws as zero-days, chained into unauthenticated remote code execution: CVE-2026-81578, an authentication bypass, and CVE-2026-82078, unsafe dynamic class loading. GTIG says security researchers and Huntress used LLMs to produce working exploits within a day of disclosure. GreyNoise reported a campaign four days after disclosure in which an attacker likely used LLMs to build and test exploits, assemble a target list and hit hundreds of organisations, expanding one foothold to domain administrator within minutes.
  • PHP. Two PHP flaws, CVE-2026-109618 and CVE-2026-109619, were found with LLM-assisted analysis. A proof-of-concept video and early patched code were published before the official patches arrived on 21 October 2026, which let both researchers and attackers compare code versions. Afterwards, hundreds of claimed but non-working proofs of concept appeared, likely LLM-generated, which GTIG says created extra verification work for defenders.

See also our earlier report on ARTEX and AI pentest frameworks.

The numbers

GTIG analysed vulnerability statistics in September. It reports that overall disclosures and exploits nearly doubled between January and August 2026, while the number of zero-day exploits rose only marginally, from an average of 8 a month in 2025 to 11 a month in 2026, before spiking to 22 in August.

Zero-day exploits per month (GTIG, as reported in its November 2026 tracker)exploits per month
Zero-day exploits per month (GTIG, as reported in its November 2026 tracker) (exploits per month)
exploits per month
2025 monthly average8
2026 monthly average11
Aug 202622

Source: Google Threat Intelligence Group, AI Threat Tracker, 16 November 2026

GTIG presents these as part of the argument that models change both the speed and the type of flaws found. The report does not claim the August spike was caused by the exploit_agent framework.

How the frameworks work

GTIG describes four open source frameworks that attackers are using or testing:

  • CAIRN runs an observe, orient, decide, act loop in which specialised agents share a "blackboard" and work from a starting point such as an IP address towards an objective. GTIG saw it used in intrusion activity against targets including the US aerospace sector.
  • PentAGI is designed to lower the skill needed to run complex campaigns. Between late August and early October 2026, GTIG tracked a threat actor using a customised platform built on it to scan a wide range of organisations in Asia.
  • CyberStrikeAI pairs a Go-based backend that acts as command-and-control with commercial and open-weight models and over 100 security tools. GTIG has seen experimentation with it.
  • HexStrike AI exposes more than 150 offensive tools to language models through the Model Context Protocol and coordinates more than 12 sub-agents. GTIG has seen experimentation with it too.

GTIG says they share core abilities: finding and exploiting low-complexity flaws, generating single-use payloads on demand, and adjusting tactics from real-time execution feedback. It says experimenting actors include financially motivated groups and suspected espionage actors with a link to the People's Republic of China, and that one such group appears to be building its own automated attack framework.

What it means

For defenders, the timing is the story. A patch used to be a signal that a fix existed; GTIG's report treats it as a starting gun for attackers who automate the comparison. That makes slow patch cycles on internet-facing software harder to justify, and it favours organisations that can deploy updates within hours.

There is a limit to the threat, which GTIG states plainly. Unrestricted autonomous execution is noisy, relies on recognisable identifiers and leaves code-level traces, so these tools trade stealth for speed. GTIG expects stealth-focused espionage groups to use them selectively, and financially motivated actors, which care less about stealth, to keep adopting them because reach can make up for lower success rates.

For developers who use AI coding assistants, the .claude/settings.json case is a reminder that agent configuration is now part of the supply chain. GTIG describes a suspected North Korean cluster (MIDNIGHT NEPTUNE in the report; SiliconANGLE writes that the group is tracked as UNC6836) using fake recruiter profiles with images carrying SynthID watermarks, ClickFix lures and GitHub tokens named after Microsoft Copilot to blend in.

The report covers more than exploitation. It says information operations have moved from experimentation to full integration of language models, including a turnkey platform designed to manage over 100,000 accounts using cloud phone farms, built in part with the Gemini API. It says LLMJacking incidents handled by Mandiant sometimes produced single-day fraudulent charges above 100,000 US dollars and, in the worst cases, cumulative abuse of 1 million US dollars. It also describes "LLM Mixers", illicit storefronts selling access to stolen model credentials, with advertised prices of 0.10 to 25 US dollars per million tokens. Google says it disabled the assets tied to the Gemini misuse it found and that DeepMind used the findings to strengthen its classifiers.

What we don't know yet

  • Who runs exploit_agent. GTIG describes Russian-language tooling and gives no name or attribution.
  • How many victims were compromised, and which flaws the framework actually exploited. The report names the Spectra plugin only as an example of code the agent was analysing.
  • Whether other defenders have seen the same framework. GTIG's evidence comes from one exposed server.
  • How representative this is. GTIG has visibility into its own telemetry and engagements, so the report cannot say how common such pipelines are across the internet.

What to watch next

Watch whether other threat intelligence teams report similar autonomous pipelines, whether vendors shorten patch-to-advisory timelines or delay publishing source diffs, and what GTIG's next quarterly tracker says about the share of exploitation that is automated. Also watch the open source frameworks themselves: their maintainers' response and how quickly detections for them improve.

Sources

Written by

audiotale.net team

The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.

Google GTIG: exploit_agent turns patches into exploits, no human involved · audiotale.net