# Google finds attackers using AI to exploit bugs with no human involved

> A framework called exploit_agent read new security patches, wrote exploits and ran them on its own, targeting French websites, Google Threat Intelligence Group says.

URL: https://claudegpt.si/news/gtig-ai-threat-tracker-autonomous-exploits  
Published: 2026-11-16T14:57:33.524Z  
Author: ClaudeGPT Si Newsroom  
Section: AI News  
Topics: Google, Cybersecurity, AI security, AI agents, Claude Code  
Publisher: ClaudeGPT Si

## Key takeaways

- Google Threat Intelligence Group says attackers used an AI framework, exploit_agent, to find and exploit patched bugs with no human in the loop, a first in the wild.
- After two PaperCut flaws were disclosed in late August, working AI-built exploits appeared within a day and a mass campaign within four days.
- A suspected North Korean group altered .claude/settings.json files so that starting Claude Code downloaded a backdoor.

Attackers have used an AI system to find and exploit security holes with no person involved, according to the [GTIG AI Threat Tracker](https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold) that Google Threat Intelligence Group published on November 16. Google says it is the first time it has seen vulnerability discovery and exploitation "fully automate[d]" by threat actors in the wild.

The finding matters for anyone who runs software on the open web. Two months earlier, GTIG had not seen fully autonomous attack pipelines in use. Now it has found one that reads each new security patch, works out what it fixed and attacks sites that haven't installed it yet.

## How the exploit_agent framework worked

GTIG found the tool, which its operator called `exploit_agent`, on an exposed staging server. It connects to several frontier and locally run language models; the report doesn't name which ones. It compares the code of popular web applications before and after each patch is announced, a method called differential analysis. When it spots a likely flaw, it drafts an exploit and runs it automatically.

The server also held scripts for what attackers do after getting in: looking around the network, guessing passwords, installing webshells (small programs that give remote control of a web server) and removing webshells left by rival attackers. Victim data on the server suggests the operator was going after French websites in media, entertainment and education. Comments in its configuration file were in Russian.

## Old bugs are the easier target

The report argues that language models make known, already-patched bugs, called n-days, more attractive than undiscovered zero-days. A patch tells the attacker exactly where to look. Users who haven't updated stay exposed.

The numbers back up the speed. When two flaws in PaperCut print servers were disclosed in late August, researchers and the security firm Huntress built working exploits within a day using language models. Four days after disclosure, GreyNoise saw a campaign that likely used models to build exploits and hit hundreds of organizations. In one case the attacker went from a first foothold to domain administrator within minutes.

There is a side effect for defenders too. After two PHP flaws were patched on October 21, hundreds of exploit demos that didn't work, likely written with AI, were published. Each one still had to be checked.

## Agent frameworks, stolen AI access and a Claude Code backdoor

The report covers more than the one tool:

- **Open-source attack agents.** GTIG saw attackers use agentic pentesting frameworks in real intrusions, including CAIRN against US aerospace targets and a customized version of PentAGI used to scan organizations in Asia from late August to early October.
- **LLMJacking.** Mandiant handled several cases in 2026 where attackers stole cloud AI credentials and ran up bills. Some single days topped $100,000 in fraudulent charges, and the worst cases reached $1 million. Stolen access is resold through "LLM Mixers," storefronts selling tokens for $0.10 to $25 per million.
- **Coding agents as a delivery route.** A suspected North Korean cluster Google calls MIDNIGHT NEPTUNE changed `.claude/settings.json` files in victims' code repositories so that starting Claude Code downloaded a backdoor called SOMBERMEME.

The last point is worth acting on now. Agent configuration files in a repository can run code when the tool starts, so they deserve the same review as any script.

## What Google did and what to watch

Google says it disabled the accounts and assets linked to misuse of its own models and that Google DeepMind used the findings to improve its classifiers. GTIG notes one weakness of these attack agents: they "currently trade stealth for velocity," leaving traffic and code traces that defenders can detect.

For teams that run web apps, the practical lesson is patch speed. The PaperCut timeline, exploits within a day and a mass campaign within four, is the window to plan for. Our explainer on [what an AI agent is](https://claudegpt.si/news/what-is-an-ai-agent) covers how these loops work.

## Sources

- [GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold](https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold), Google Cloud
- [Google says an AI framework turned security patches into exploits with no human in the loop](https://siliconangle.com/2026/11/16/google-says-an-ai-framework-turned-security-patches-into-exploits-with-no-human-in-the-loop/), SiliconANGLE
