Google's Threat Intelligence Group says it found, for the first time, attackers using an AI framework that compares code before and after a patch, writes an exploit and runs it without a human. Its November 16 report also names CAIRN, CyberStrikeAI and PentAGI as agent frameworks now used in real intrusions.
Google’s Threat Intelligence Group (GTIG) says attackers have fully automated the step from security patch to working exploit. In its AI Threat Tracker report published November 16, GTIG and Mandiant say that in the past quarter they saw, for the first time, threat actors using AI frameworks to find and exploit vulnerabilities “without a human in the loop.”
If your team patches internet-facing software on a monthly cycle, that cycle is now the risk. Patch public web apps, VPNs and plugins within days of a fix, and turn on automatic updates where the vendor offers them.
What Google found
- A patch-to-exploit pipeline. GTIG found an exposed staging server running a multi-agent framework it calls exploit_agent. It watches popular web application software, compares the code before and after each patch, works out the flaw, drafts an exploit and runs it automatically. Victim data on the server pointed at French media, entertainment and education websites.
- Agent frameworks in live intrusions. Besides ARTEX, which the FBI warned about last week, GTIG says attackers now use open source agent frameworks named CAIRN, CyberStrikeAI and PentAGI. It saw CAIRN in intrusions against US aerospace targets and a customized PentAGI used to scan organizations in Asia between late August and early October.
- More exploits overall. GTIG says vulnerability disclosures and exploits nearly doubled between January and August 2026. Zero-day exploits rose more slowly, from an average of 8 a month in 2025 to 11 in 2026, then spiked to 22 in August.
- Fake exploits too. After the October PHP patches, hundreds of AI-generated proof-of-concept exploits were published that did not work, which added verification work for defenders.
- Coding assistants as a target. Espionage groups are tampering with the settings of AI coding assistants so a backdoor drops when the assistant runs on a developer’s machine, the report says.
There is a limit. GTIG notes these frameworks “trade stealth for velocity”: they generate noisy traffic and leave code traces that make detection more likely. The Register and BleepingComputer both led their coverage with the no-human exploit finding.
Monthly patching is too slow for internet-facing software
This is our analysis. The time between a vendor’s fix and a working attack used to give small teams a few weeks of cover. An agent that reads the diff and writes the exploit removes most of that. For a remote company, the exposed surface is usually short: the website and its plugins, a VPN or remote-access gateway, a NAS, and any self-hosted tool reachable from the internet.
Make a list of those systems this week, and give each one an owner and an automatic update setting where it exists. Keep production credentials off developer laptops that run AI coding assistants, and review assistant configuration files in your repositories the way you review code.
Related: the FBI’s warning about the ARTEX AI hacking tool, the LLMjacking schemes that run up AI bills with stolen cloud keys, and the backdoored Claude Code extension spread through Cursor.
Related, November 16: Anthropic is cutting its disclosure deadline for severe bugs to 30 days, so vendors get less time to ship a fix.
