Google spots first fully autonomous AI turning security patches into live attacks

Google's Threat Intelligence Group found an AI system called exploit_agent sitting on an exposed server. It compares software before and after security updates, writes attack code on its own and then uses it right away. Google says this is the first fully autonomous AI attack it has seen in the wild.

Industry Nov 16, 2026 7 min read Source: Techmeme
Illustrative image, AI-generated

What's going on

The Google Threat Intelligence Group (GTIG), the part of Google that analyzes cyber threats, published its quarterly report on how attackers misuse AI on November 16, 2026. It's titled "Agentic Threat Activity Crosses a Significant Threshold". At its center is a finding Google calls a first. For the first time, researchers watched attackers run an AI framework with no human in the loop at all. The system analyzed security updates, built exploits from them (code written to take advantage of a specific vulnerability) and automatically used them against real targets.

The framework is called exploit_agent. It turned up on an unprotected staging server, a machine the attacker used for preparation that had been left open to the internet by mistake. SiliconANGLE and the Spanish-language outlet DiarioBitcoin are among those reporting on it.

How exploit_agent works

The idea is simple. When a software vendor fixes a security hole, it changes specific parts of the code. If you compare the version before the update with the one after, you can see what was fixed and work out what the bug was. This is called patch diffing, and until now it has been slow manual work for experienced security researchers.

According to GTIG, exploit_agent automated this step completely. The system keeps comparing the code of popular web applications before and after every patch announcement. When it spots a likely vulnerability, it writes an exploit and runs it immediately. To do this, it was hooked up to several large language models: frontier models, meaning the most capable commercial AI models, and models the operator ran on their own hardware. According to DiarioBitcoin, the report doesn't say which models were involved.

DiarioBitcoin explains what's new here. The system didn't just come up with hypotheses or suggest tests for a human to sign off on. It carried out the whole chain itself, from analysis to attack. That's what sets this case apart from the usual pattern so far, where attackers ask a chatbot questions and check the answers themselves.

What else was on the server

On the same server, researchers found stolen victim data and several scripts for use after a break-in. These could map internal networks, crack passwords by brute force (simply trying one combination after another) and install webshells. Webshells are small programs that keep a backdoor open on a web server for an attacker. One script even removed webshells left behind by rival attackers. The operator apparently wanted the hijacked servers all to themselves.

According to victim telemetry, the operator was mostly interested in French websites in media, entertainment and education. The configuration file contains a comment in Russian telling whoever sets it up to enter the keys "right here, not in the chat". DiarioBitcoin points out that Russian-language tools alone don't tell you who's behind an attack. The report doesn't attribute the activity to any state. It also doesn't name the victims, say how much data was stolen or identify the software that was analyzed.

Why attackers go after known vulnerabilities

In its report two months earlier, GTIG still said it hadn't seen any fully autonomous attack chains against real targets. That changed in the last quarter. Google expects attackers to point language models at n-day vulnerabilities rather than hunt for zero-days. N-day vulnerabilities are publicly known and already patched, but not yet fixed everywhere. Zero-days are flaws the vendor doesn't know about yet. The reasoning: a patched vulnerability comes with a limited, well-defined body of material, namely the changed code. That's easier for an AI to work through than an open-ended search for unknown bugs. According to DiarioBitcoin, though, Google also mentions the first known case of an attacker using an AI-developed zero-day exploit in a real attack.

The report uses two vulnerabilities in PaperCut print management servers to show how fast the gap between disclosure and the first working attack is shrinking. They were chained together and exploited as zero-days in late August. Security researchers and the company Huntress Labs used language models to build working exploits within a day of disclosure. A campaign that GreyNoise Intelligence spotted four days after disclosure is also said to have built its exploits with language models, and it hit hundreds of organizations. In one case, GreyNoise says, the attackers gained domain admin rights, meaning control over the entire corporate network, within minutes.

More findings: pentest frameworks, fake profiles, tampered Claude Code

The report also describes how attackers abuse freely available pentest frameworks with AI agents. These tools are meant for penetration testing, where professionals attack a company's systems on its behalf to find weaknesses. The multi-agent framework CAIRN was seen during an intrusion in the US aerospace industry. Another actor used a customized platform based on PentAGI from late August to early October to scan many organizations in Asia for vulnerabilities. DiarioBitcoin also names ARTEX, CyberStrikeAI and HexStrike AI. According to GTIG, both financially motivated criminals and suspected espionage groups are experimenting with these tools, some of them with ties to China. One small consolation: because these frameworks work at full throttle with no regard for stealth, they leave lots of traces and have so far been fairly easy to detect.

In influence campaigns, too, AI use now goes far beyond writing posts. A commercial provider used the Gemini API to build a turnkey platform meant to run more than 100,000 social media accounts through cloud phone farms. Another platform targeted political sentiment in Europe and the Middle East. Its prompts deliberately asked for camera flaws so the generated photos would look like real smartphone shots.

The case of UNC6836, a suspected North Korean group, is especially concerning. It approached crypto companies in North America and Europe through fake recruiter profiles. The profile photos carried Google's SynthID watermark, an invisible marker for AI-generated images, which means they were made with Google tools. Later, the group tampered with .claude/settings.json files in internal code repositories. As a result, Anthropic's coding assistant Claude Code downloaded the SOMBERMEME backdoor every time it started. Google says it has disabled the accounts and resources linked to the Gemini abuse it found. Google DeepMind used the findings to improve its classifiers, the filters meant to detect abuse automatically.

The bigger picture

The discovery of exploit_agent confirms what security researchers have been warning about for a while: the time between a security update and the first attack on unpatched systems is shrinking. If an AI analyzes updates around the clock and turns them into attacks automatically, taking weeks or even just days to install a patch is no longer good enough. If you run web applications, that means installing updates immediately and watching for suspicious access. A patch only protects you once it's installed, and publishing it also tells attackers exactly where to look.

But the case doesn't show that any AI can exploit any vulnerability on its own, or that humans have dropped out of attack campaigns. DiarioBitcoin stresses this too. The report describes a single framework on a single exposed server. It doesn't say how many victims there were, which models were used or how successful the system was overall. Google is also a vendor of AI models and security services itself, and outsiders can't verify its case numbers. It stands out that several cases directly involve tools from major AI vendors: the Gemini API, SynthID-marked images and Claude Code. The vendors' safeguards often seem to kick in only after the fact.

Sources

#cybersecurity#google#ai agents#exploits#misuse
Read the original source

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.