Skip to content
AI Security Wire

Published

- 5 min read

By

GTIG: First Fully Automated Exploit Pipeline Found Running Without a Human

img of GTIG: First Fully Automated Exploit Pipeline Found Running Without a Human

Google Threat Intelligence Group (GTIG) published its latest AI Threat Tracker today, and one finding stands out from the usual “attackers use chatbots” summary. GTIG says that in the past quarter it identified, for the first time, threat actors using LLM frameworks to automate vulnerability discovery and exploitation with no human in the loop.

The exploit_agent server

GTIG found an exposed staging server hosting a multi-agent framework it calls exploit_agent. Per the report, it connects to several frontier and local LLMs and runs continuous differential analysis of popular web application software. It compares code before and after a patch announcement, analyzes the difference for vulnerabilities, drafts exploits and executes them automatically.

The server also held post-compromise scripts for internal reconnaissance, brute-forcing, installing webshells and removing other attackers’ webshells. GTIG says victim telemetry on the box points to an operator interested in French media, entertainment and education sites. Comments in the environment file were in Russian. The report names no operator.

One detail is worth a pause: this is a single exposed server, observed by one vendor. Our assessment: it shows such a pipeline exists. GTIG says the framework “is set up to” operate this way and that the server held data exfiltrated from victims, but the server does not tell us how many actors run something similar.

Why n-days are the target

GTIG’s reasoning is simple. A patched n-day gives an attacker a finite pile of material: two code versions, release notes, earlier research and proof-of-concept code. Hunting a zero-day means casting a wider net. GTIG says threat actors may find it more efficient to use LLMs to weaponize patched n-days.

Our assessment: GTIG’s own vulnerability statistics are consistent with that. It says overall disclosures and exploits nearly doubled between January and August 2026, while zero-day exploits rose only from a 2025 average of 8 a month to 11 a month in 2026, with a spike of 22 in August.

GTIG also cites the PaperCut NG/MF campaign. GreyNoise’s own write-up says an operator working from 31 August 2026 used hundreds of agents, built on OpenAI’s Codex harness with a DeepSeek model, to compromise at least 440 PaperCut instances at 395 organizations in 48 countries. GreyNoise reports first remote code execution on a real target in just under four hours from an empty workspace, and domain admin at 12 of those victims. Those are GreyNoise’s figures. The same report says Cloudflare’s WAF stopped the operator against at least one host, a reminder that ordinary hardening still bites.

GTIG adds a side effect on the defender side: exploits claimed to work that do not, which GTIG says were likely generated with LLM help. After the PHP patches of 21 October 2026, GTIG says hundreds of non-functional proofs of concept appeared, and verifying them took real time.

Open-source offensive frameworks in the wild

GTIG says it has independently seen CAIRN and PentAGI deployed in live activity, and experimentation with CyberStrikeAI and HexStrike AI. Two examples from the report:

  • CAIRN was used in intrusion activity against targets that include the US aerospace sector.
  • Between late August and early October 2026, GTIG tracked an actor running a customized orchestration platform built on open-source PentAGI for reconnaissance and vulnerability scanning against organizations in Asia.

GTIG’s caution is useful: these tools currently trade stealth for velocity. Unrestricted autonomous execution generates forensic traffic and leaves recognizable identifiers. Cybercrime actors who care less about stealth will keep using them, in GTIG’s high-confidence assessment.

The agent config file as an attack path

The most practical item for defenders is a suspected DPRK cluster GTIG tracks as MIDNIGHT NEPTUNE. In intrusions at cryptocurrency organizations in North America and Europe, GTIG says the actor used fake recruiter profiles (profile images carried SynthID watermarks, which GTIG treats as confirmation of LLM-synthesized identities) and ClickFix lures for initial access.

After that, the actor reportedly pivoted into internal code repositories and modified .claude/settings.json so that Claude Code would trigger the download of the SOMBERMEME backdoor when users started it. It also created GitHub personal access tokens named after Microsoft Copilot to blend in. GTIG attributes this to a “suspected” cluster, so treat the attribution as provisional.

Smaller items

  • LLMjacking: Mandiant says some incidents hit single-day fraudulent charges above $100,000 and, in the worst cases, cumulative abuse of $1 million before remediation. GTIG also tracks “LLM Mixer” storefronts selling stolen model access at roughly $0.10 to $25 per million tokens.
  • Influence operations: GTIG assesses with high confidence that IO actors have moved from experimenting with LLMs to full operational integration.

What to do with this

Our assessment: the patch-to-exploit window is the number to watch, so measure how long internet-facing software stays unpatched after an advisory. Add agent configuration files to code review and change monitoring. And expect noisy, fast intrusions for now, so volumetric and behavioral detections are worth tuning. GTIG itself anticipates reduced human involvement and “increased stealth and sophistication” as the tools and models improve, so don’t bank on the noise lasting.

References

Frequently Asked Questions

What is exploit_agent?
It is the name GTIG gave to a multi-agent vulnerability research and weaponization framework it found on an exposed staging server. According to GTIG, it compares code before and after patch announcements, drafts exploits, and runs them automatically. GTIG calls this the first time it has seen a threat actor automate discovery and exploitation without a human in the loop.
How fast can attackers turn a patch into an exploit now?
GTIG says LLMs shrink the window between patch disclosure and weaponized exploitation. In a separate PaperCut campaign, GreyNoise reports that the operator went from an empty workspace to a first real-world remote code execution in just under four hours. Both are vendor-reported observations, not a universal benchmark.
What should defenders do about agent config files like .claude/settings.json?
GTIG describes a suspected DPRK cluster editing that file in internal repositories so Claude Code would fetch a backdoor on startup. Our assessment: treat agent configuration and hook files as executable code, put them under code review and change alerting, and check them after any repository compromise.