Google Threat Intelligence Group (GTIG) says it has seen attackers fully automate both finding and exploiting vulnerabilities, with no human in the loop, for the first time. According to CyberScoop and The Register, the case appears in GTIG's quarterly AI threat tracker, published on November 16. GTIG and Mandiant found the system, called exploit_agent, on an exposed staging server. It watched popular web software for security patches, worked out what each fix revealed, wrote an exploit for it and then ran that exploit.
Why it matters: most organisations plan their patching on the assumption that an attacker needs days or weeks to turn a patch into a reliable exploit. A system that does this automatically and around the clock shortens that window for everyone, not just for high-value targets.
What Google found
| Detail | What was reported |
|---|---|
| Tool | "exploit_agent", a multi-agent vulnerability research and weaponisation system |
| Models | Mixes frontier models with locally hosted ones; the models aren't named in the reports |
| Method | Continuously compares code before and after patches in "popular web application software", writes exploits, runs them automatically |
| Other tooling on the server | Scripts for internal reconnaissance, password brute-forcing, installing webshells and removing other attackers' webshells |
| Operator | Suspected Russian-speaking, based on Russian-language code comments and a config file that told users where to put API keys for "auto-exploitation" |
| Targets | Victim data on the server pointed to French websites in media, entertainment and education |
None of the four reports we read names a CVE, a vendor, a victim count or how quickly the system moved from patch to exploit. Google also didn't name a group or a state. Techzine notes that the disclosure focuses on techniques, not on who was behind them.
The wider picture in the tracker
CyberScoop reports two more data points from Google:
- More zero-days. Google counted 22 exploited zero-days in August 2026 alone. That compares with about 11 a month earlier in 2026 and an average of eight a month in 2025.
- Many fake exploits. After vendors patched two PHP vulnerabilities on October 21, hundreds of claimed exploits appeared within days. Most didn't work and were likely AI-generated.
The Register lists other offensive AI frameworks that attackers are using: ARTEX, CAIRN, CyberStrikeAI, HexStrike AI and PentAGI. GTIG saw PentAGI used through proxy services against aerospace targets. It also says financially motivated criminals and suspected PRC-linked espionage groups are both testing these tools. The Decoder says Google rates the current activity at "Level 5" on its own maturity scale, meaning AI carries out several attack stages with little human involvement. It also says Google expects more capable attacks within months.
How much weight to give it
- The "first" is Google's assessment. CyberScoop reports that Google says it believes this is the first fully automated, end-to-end intrusion. No one outside Google has confirmed it.
- Attribution is circumstantial. It rests on the language in the code and config files and on the targeting data.
- The attacks were loud. The Decoder and CyberScoop say these tools trade stealth for speed and leave forensic traces. Techzine adds that the system had no way of cleaning up after itself.
- We couldn't read Google's own post. This article relies on the four outlets listed below, which all summarise it. Details such as how the maturity levels are defined come from their accounts.
What defenders should do
These steps are our reading of what the reports describe. They are not Google's guidance.
- Shorten the patch window for internet-facing web apps. For a CMS, a PHP application or a plugin exposed to the internet, assume that a patch release starts the clock for attackers. Aim to deploy within days.
- Put compensating controls in place on release day. If you can't patch quickly, use WAF rules, restrict access to admin paths, or take the app offline briefly when a security fix lands.
- Hunt for the noise. Automated campaigns probe heavily. Look for bursts of requests that match newly patched code paths, brute-force attempts, and new or deleted webshells in web roots. Deleted webshells matter because this tooling removed other attackers' shells.
- Treat a compromised web server as a credential problem. The tooling included reconnaissance and brute-force scripts, so rotate the secrets that server can reach.
- Watch AI-generated proof-of-concept exploits. Most of the PHP "exploits" Google saw didn't work, so test before you trust a public PoC. That includes testing it as proof that you are safe.
Google expects these frameworks to keep spreading. The OpenAI agent incident at Hugging Face this summer led to new oversight arrangements. This case is different: the operators were human, and they deliberately pointed their agents at victims.
Sources
- CyberScoop — Google says Russian-speaking hackers ran the first fully automated AI cyberattack cyberscoop.com
- The Register — Google: AI agent turned a security patch into a working exploit with no human help theregister.com
- The Decoder — Google warns that fully automated cyberattacks have arrived and will get worse fast the-decoder.com
- Techzine — Russians move away from human-in-the-loop for fully autonomous cyberattacks techzine.eu