For years, defenders have relied on a rough rule: when a vendor ships a security patch, you have some time before attackers work out what it fixed and turn that into a working exploit. Google's threat researchers say that time is shrinking, because some of the work is now being done by AI agents.
In its latest AI Threat Tracker, published November 16, Google Threat Intelligence Group (GTIG) says adversary use of AI has moved past chatbots that help draft phishing emails. Attackers are now running autonomous, multi-agent frameworks. In one case GTIG describes, a suspected Russian-speaking operator used an AI agent to go from a vendor's security patch to working exploits without a human stepping in. The report is credited to Aurora Blum, Kelli Vanderlee, Jess Xia and David Mainor.
The exploit_agent case
The most striking finding concerns an operator The Register describes as suspected to be Russian-speaking. According to The Register's reporting on the research, GTIG found an exposed staging server hosting a framework called exploit_agent. 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 Register adds that the server also held data stolen from victim organizations.
The framework did three things, The Register reports:
- Continuous differential analysis of popular web application software. "Patch diffing" means comparing a program's code before and after a security update. The changed lines point to where the bug was. Here the comparison was automated and continuous, followed by automatic drafting of exploits.
- Automatic exploit execution against targets, with no operator approving each step.
- Post-compromise scripts for internal reconnaissance, brute-forcing, and installing webshells (and removing other attackers' webshells). Webshells are small backdoor scripts that give an attacker continued remote control of a web server.
GTIG says victim telemetry on the server suggests the operator was interested in French websites associated with the media and entertainment industries, as well as the education sector.
The research is about n-day vulnerabilities: flaws that have been publicly disclosed and patched but are still present on systems that haven't applied the update. A zero-day is unknown to the vendor. An n-day is a race between the people applying the patch and the people reverse engineering it. Automating the reverse-engineering side gives attackers a head start in that race.
The Hacker News, which also received the GTIG report, says the report cites PaperCut NG/MF flaws as a case where attackers rapidly weaponized newly disclosed vulnerabilities. GTIG's own report says that in late August 2026 threat actors exploited two chained PaperCut print-management vulnerabilities, CVE-2026-81578 and CVE-2026-82078, as zero-days. It adds that security researchers and Huntress used LLMs to generate working exploits within a day of disclosure, and that Greynoise reported a campaign four days after disclosure in which an actor likely used LLMs to build exploits and attack hundreds of organizations.
A crowded field of offensive agent frameworks
exploit_agent is not the only tool in the report. GTIG names several autonomous multi-agent frameworks it has seen adversaries adopt: ARTEX, CAIRN, CyberStrikeAI and PentAGI. GTIG has also seen experimentation with HexStrike AI, and the report also names that the report also names Agent Zero and ZeroHollow, which it describes as an inference service for exploit development and LLM routing.
GTIG describes several of these as open source penetration-testing frameworks, the kind of tooling security teams use to test their own systems. Software that can automate an authorized assessment can also automate an unauthorized one. GTIG's main point is that the automation now reaches all the way from finding the bug to using it.
The rest of the threat picture
The report covers several other trends.
LLMjacking for resale. LLMjacking is the theft of access to commercial AI models, usually through stolen API keys or cloud credentials. GTIG says stolen LLM credentials and cloud resources are being resold through underground proxy services, which it also calls LLM mixers. The Hacker News names two such services, APIMart and JetNight, and says access is sold at discounted prices.
Phishing kits with built-in AI. According to The Hacker News, a Chinese-language phishing platform called Jieyou includes an LLM chatbot assistant that lets customers configure phishing panels and modify templates using natural language. Phishing-as-a-service platforms sell ready-made kits to less skilled criminals, and a built-in AI assistant makes it easier for those customers to write convincing, targeted messages.
Influence operations. GTIG reports that LLMs continue to be used in influence operations, including a long-running provider it has tracked since 2019 that uses them to raise the quality and volume of its output.
Espionage through developer tools. The report flags supply-chain tricks by espionage actors, including manipulating autonomous coding assistant configurations to drop backdoors during local execution. Many developers give these assistants broad access to source code, terminals and credentials, so a tampered configuration can quietly change what the assistant does or what it can reach.
What It Means for Defenders
Who is affected: Anyone running internet-facing web applications, especially popular, widely deployed software. Automated patch diffing pays off most when the same code runs on many servers. GTIG's telemetry pointed to media, entertainment and education sites in France, but nothing about the technique is tied to one country or sector. Schools and offices running PaperCut NG/MF are a named example.
What to patch and check:
- Shorten your patch window for internet-facing software. GTIG's PaperCut examples show working exploits within a day of disclosure, so assume a public patch can become a working exploit within days, not weeks. Put exposed web apps, VPNs, print servers and admin panels at the front of the queue.
- PaperCut NG/MF: GTIG says two chained PaperCut vulnerabilities (CVE-2026-81578 and CVE-2026-82078) were exploited as zero-days in late August 2026. The sources do not list specific fixed versions. Confirm you are on the latest release from PaperCut's own advisories, and check whether the server is reachable from the internet. If it doesn't need to be, take it off.
- Hunt for post-compromise activity. exploit_agent's follow-on scripts did recon, brute-forcing and webshell deployment. Look for unexpected script files in web roots, bursts of failed logins, and new outbound connections from web servers. If a system went unpatched for a while after a fix shipped, investigate it as possibly compromised. Patching alone won't remove an attacker who is already in.
- Rotate and lock down AI API keys. Given the resale market for stolen LLM access, treat model API keys and cloud AI credentials like any other high-value secret. Keep them out of source code, scope them tightly, set spending alerts, and rotate any that may have been exposed.
- Audit AI coding assistant configurations. Because espionage actors are tampering with these settings, review what your developers' assistants are allowed to access and run. Watch for unexpected changes to their config files, and treat those files as security-relevant code under review.
- Train against tailored phishing. Tools like Jieyou's assistant mean lures will be more polished and more specific to the target. Phishing-resistant MFA does more here than awareness training alone.
For small businesses and CMMC-bound contractors, this mostly reinforces controls you already have. Timely flaw remediation, an accurate inventory of internet-facing assets, and monitoring for unauthorized changes are baseline requirements. What GTIG's report changes is how quickly they need to happen.