# exploit_agent: patch to exploit with no human in the loop

> Patch to exploit now runs end to end without a person, according to Google Threat Intelligence Group's November AI Threat Tracker.

- **Severity:** Info: Research, tools, evals or lab claims with no direct exposure. Context, not an alert.
- **Category:** Incident, Attack
- **Published by:** Google Threat Intelligence Group
- **Disclosed:** 2026-11-16
- **Affects:** Spectra (Ultimate Addons for Gutenberg), WordPress, Claude Code
- **Research by:** Aurora Blum, Kelli Vanderlee, Jess Xia, David Mainor
- **Primary source:** https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold
- **Page:** https://theinjection.dev/items/gtig-exploit-agent-autonomous-exploitation/

exploit_agent is a multi-agent framework that compares software before and after a patch, works out the flaw, writes an exploit and runs it without a person. Google Threat Intelligence Group found it on an exposed staging server.

## What happened

Google Threat Intelligence Group published its AI Threat Tracker on 16 November 2026 and described exploit_agent, a framework it found on an exposed staging server. GTIG says this is the first time it has seen threat actors fully automate vulnerability discovery and exploitation without a human in the loop. The server also held exfiltrated victim data and scripts for internal reconnaissance, brute-forcing and installing webshells.

## How it works

exploit_agent runs continuously against popular web application software and connects to several frontier and local models. The framework performs differential analysis, comparing code before and after a patch announcement to work out which flaw the patch closed, then drafts an exploit and executes it. GTIG's Figure 4 shows the framework reading an access control check in the Spectra plugin for WordPress, version 2.19.25, that guards the /wp/v2/spectra-popup REST route.

## Who is affected

Internet-facing web applications carrying recently patched n-day flaws are the exposure here, because the gap between a vendor patch and a working exploit keeps narrowing. Victim telemetry on the server pointed at French websites in the media, entertainment and education sectors. GTIG separately reports that a suspected MIDNIGHT NEPTUNE cluster edited .claude/settings.json files so that starting Claude Code downloaded the SOMBERMEME backdoor.

## What to do

Shorten patch windows, which is GTIG's main recommendation, because n-day exploitation is accelerating. Review .claude/settings.json and other agent configuration files in your repositories for commands nobody on the team added.

## Quick facts

- **Framework:** exploit_agent
- **Found by:** Google Threat Intelligence Group
- **Where:** Exposed staging server
- **Method:** Differential analysis of pre- and post-patch code
- **Example target:** Spectra for WordPress 2.19.25
- **Targeting:** French media, entertainment and education sites
- **Published:** 2026-11-16

## Questions

### Does exploit_agent target a specific vulnerability I should patch?

Google Threat Intelligence Group names no CVE for exploit_agent. The report shows the framework reading an access control check in the Spectra plugin for WordPress version 2.19.25, as an example of the code it analyzes. GTIG's guidance is general rather than product specific: shorten patch windows, because the gap between a vendor patch and a working n-day exploit keeps narrowing.

### How did GTIG find exploit_agent?

GTIG found exploit_agent on a staging server the operator left exposed. Alongside the framework, the server held exfiltrated victim data and scripts for internal reconnaissance, brute-forcing, installing webshells and removing other attackers' webshells. A configuration file holding several LLM API keys carried Russian-language comments that called the setup an exploit_agent pilot. GTIG did not name the operator.

### Is Claude Code involved in the activity GTIG describes?

Claude Code appears in a separate part of the GTIG report, not in exploit_agent. GTIG says a suspected MIDNIGHT NEPTUNE cluster modified .claude/settings.json configuration files so that starting Claude Code automatically downloaded the SOMBERMEME backdoor. The same actor generated GitHub personal access tokens named after Microsoft Copilot, which made its lateral movement look like routine tooling.

## Sources

1. [GTIG AI Threat Tracker](https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold)
2. [Mandiant Hardening Guide](https://cloud.google.com/blog/topics/threat-intelligence/artex-and-friends-enter-agentic-offensive-frameworks)
3. [GTIG LLMjacking Report](https://cloud.google.com/blog/topics/threat-intelligence/llmjacking-monetizing-compromised-cloud-models)
