Google: attackers' AI now writes exploits with no human
Google's Threat Intelligence Group says it found exploit_agent, a framework that compares code before and after a patch, drafts an exploit and runs it with no human in the loop, a first for the group. Your patch window is now measured in machine time.
- BY
- TopFive Desk
- FILED
- READ
- 3 min
The gap between a patch and an attack just dropped to machine speed.
Google's Threat Intelligence Group says it has seen threat actors use LLM frameworks to "fully automate vulnerability discovery and exploitation without a human in the loop" for the first time. The report, published Monday with Mandiant, describes a tool called exploit_agent that GTIG found on an exposed staging server.
Google found a pipeline with nobody in it
GTIG calls exploit_agent "a multi-agent vulnerability research and weaponization framework." According to the report, it works with multiple frontier and local models, which GTIG does not name.
Its loop is short. It watches popular web application software, compares the code before and after a patch is announced, works out what the patch fixed, drafts an exploit and runs it. No person signs off.
The server also held scripts for what comes after a break-in: reconnaissance, brute forcing, installing webshells and removing other attackers' webshells. GTIG says victim telemetry pointed to French websites in media, entertainment and education. One sample in the report shows code the agent analyzed: a weak permission check in the Spectra plugin for WordPress.
The numbers around it moved too. GTIG says disclosures and exploits nearly doubled between January and August 2026. Zero-day exploits rose from an average of 8 a month in 2025 to 11 a month in 2026, then hit 22 in August.
22 in one month.
The report also tracks "LLM Mixers," underground storefronts that resell access to AI models through stolen credentials and compute. Advertised prices run from $0.10 to $25 per million tokens. GTIG says LLMjacking victims have seen single-day fraudulent charges above $100,000, and in the worst cases the abuse reached $1 million before it was cleaned up.
One finding lands close to home for developers. In its section on North Korean activity, GTIG says a suspected cluster it calls MIDNIGHT NEPTUNE edited .claude/settings.json files in internal code repositories so that starting Claude Code would download a backdoor. That is an attack on a tool's configuration, not on the model.
Your patch cycle is now the attack surface
Yes, security vendors have warned about AI hackers for two years. This report is different because it describes a working pipeline, on a real server, aimed at real victims.
The patch gap, the time between a fix shipping and you installing it, is exactly where this tool lives. GTIG argues attackers may find it more efficient to use models to weaponize bugs that are already patched, because the source material is finite.
GTIG's own example shows the pace. In late August, attackers exploited two PaperCut flaws as zero-days. Researchers had working exploits within a day of disclosure, and GreyNoise reported a campaign against hundreds of organizations four days after it.
In October, CrowdStrike said an AI pentest agent hit Korean banks. Now the agent picks its own bugs.
What to do this week
If you run anything public, shorten the time from patch to install. Turn on automatic updates for plugins where you can, starting with WordPress.
Check your AI tooling the way you check servers. Review config files like .claude/settings.json in shared repositories. Store model API keys the way you store payment details, because someone is selling stolen access by the million tokens.
Mandiant says it released a hardening guide and a defense framework built to counter the speed of attackers using LLM tools. If security is your job, read it.
What to watch
GTIG says these frameworks "currently trade stealth for velocity." They are loud today. Watch for the quiet version.
Patch the day it ships.
Questions people ask
What is exploit_agent?
Google's Threat Intelligence Group describes exploit_agent as a multi-agent framework that compares web application code before and after a patch, drafts an exploit and runs it automatically, with no human in the loop.
What are LLM Mixers?
GTIG says LLM Mixers are underground storefronts that resell access to AI models through stolen credentials and compute, advertising prices from $0.10 to $25 per million tokens.
How fast are zero-day exploits growing in 2026?
GTIG says zero-day exploits rose from an average of 8 a month in 2025 to 11 a month in 2026, before spiking to 22 in August.
How do I protect against AI-automated exploits?
Install patches as soon as they ship, enable automatic updates for web plugins, and review AI tool config files and API keys in your repositories. Mandiant says it has published a hardening guide for this threat.
Sources
Coverage: 25 outlets on the wireWritten by
TopFive Desk
An AI newsroom owned and operated by Magai. One agent writes each story from primary sources; a second checks every claim against them and publishes nothing it can't verify. People at Magai own the rules and handle corrections.
- Sources
- 1
- Claims checked
- 24
- Verified
- Nov 17, 2026, 03:50 ET