Infrastructure & Security · Nov 18, 2026
No human from patch release to attack: Google confirms first autonomous AI-driven exploitation
An AI system read the differences in security patches, wrote exploit code from them and carried out the attacks itself. The gap between a fix's release and its first exploitation is shrinking from days to hours, and defenders now have to apply patches faster.
Seiichi Tanaka · Editor-in-Chief

Key points
- GTIG found a system in which AI wrote exploit code from patch diffs and then ran it. GTIG says this is the first time it has seen attacks automated with no human in the loop
- Publishing a patch now acts as a signal to attack, and the window to respond is shrinking from days to hours. By GTIG's count, zero-day exploitation averaged 11 cases a month in 2026 and reached 22 in August
- Companies with internet-facing servers need to review their target times for applying critical patches and their agreements with outsourced maintenance providers
Google's threat intelligence team, GTIG (Google Threat Intelligence Group), said in a report published Nov. 16–17 that it had confirmed a new kind of case for the first time. According to the report, attackers used a system built around large language models (LLMs) to automate the whole process, from finding software flaws to carrying out attacks, with no human involved. The time between a software vendor releasing a fix (a patch) and that flaw being used in attacks is shrinking from days to hours. Batching patches to apply next week no longer works as a default approach.
For Japanese companies: IT staff at companies with their own websites or internet-facing servers (media, e-commerce, schools and education businesses, among others) should check with their maintenance contractors this week on the target time for applying critical patches. If the contract allows several days or more, they should agree on a procedure and pricing for applying emergency patches within a short time.
AI handled everything from patch diff to exploit code
The automated attack system GTIG found was called "exploit_agent" and sat on an internet-facing server. According to GTIG's report, exploit_agent compares a program's code before and after a patch to work out which flaw was fixed. It then writes exploit code targeting that flaw and launches real attacks without any human checking the code. The targets were media, entertainment and education websites in France. Data stolen from victims was also found on the server.
In the past quarter, GTIG identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time.
Here is what GTIG means by "for the first time." Attackers were not having AI write a draft for a human to finish. They had handed the whole attack process to machines, and this was found in a real attack, not an experiment. The exploit_agent case shows that AI has moved from helping with attacks to running them.
GTIG's report also said that CAIRN and PentAGI, two open-source offensive tools, had been used in real intrusions. AI-powered attack tools are now available to anyone and are being built into real attacks.
Patch releases become a signal to attack
GTIG explains why the time from patch release to attack is shrinking. When AI capabilities are chained together, reading the diff, writing exploit code and running it happen as one continuous process.
By chaining these capabilities together, threat actors can dramatically reduce the window between patch disclosure and weaponized exploitation.
A patch also works as a map showing where the flaw was. In the past, it often took human experts a while to analyze the diff and build working exploit code, and users could apply the fix in the meantime. If machines do that work instead, the race against attackers begins the moment a patch is released. Every hour before users apply the fix becomes a chance to attack.
Researchers had already shown in experiments that AI can turn patches into exploit code. In October, the research group Epoch AI reported that it had used GLM 5.3, a freely available model from China's Z.ai, to build working exploit code from published Firefox patches. It succeeded on all six versions it tested. The fastest took about 12 hours, and each version cost about $1,000 (about ¥160,000). GTIG's new report shows that what Epoch AI did in its experiment is now happening in real attacks.
By GTIG's count, flaws exploited before a fix is available (zero-days) averaged 8 a month in 2025. In 2026 the average rose to 11 a month, and in August it reached 22.
A flood of broken exploit code takes up defenders' time
AI-driven attack automation is also putting a different burden on defenders. GTIG says that because it is now easier to get into attacking, large numbers of LLM-written exploits are circulating that are claimed to work but don't.
Simultaneously, the lower barrier to entry has fueled a wave of LLM-generated exploits that are claimed to work but don't, creating additional verification work for defenders.
Checking whether circulating exploit code is real falls to security staff and to companies that distribute vulnerability information. The more fake exploits there are, the easier it is for real alerts to get lost. Working out whether a threat affects your own company also takes time.
AI development tools themselves were targeted
GTIG's report also described attackers going after AI tools themselves. A group believed to be linked to North Korea modified the configuration files of Claude Code, a tool used by developers, so that a backdoor (a mechanism for remote control) ran every time the tool started. In another case, stolen login credentials for AI services were abused to run up usage charges of more than $100,000 (about ¥15.9 million) a day, reaching $1 million (about ¥159 million) in total.
Companies that use AI development tools such as Claude Code internally should have their IT departments add two checks: whether configuration files on developers' machines have been changed without authorization, and whether spending on AI services has suddenly risen.
Patching moves from "days" to "hours"
If the time from patch release to attack is shrinking, defenders have no choice but to shorten the time it takes them to patch. Jason Elrod, chief information security officer (CISO) at U.S. healthcare provider MultiCare Health System, said in U.S. outlet TechTarget's coverage of the GTIG report that traditional patching cycles no longer hold up.
Traditional patching cycles are becoming obsolete, and some fixes will need to happen in hours rather than days or weeks
Neil Wyler, vice president of defensive services at security firm Coalfire, takes the same view. As attack timelines shrink, he said, the time from spotting an anomaly to containing the damage has to shrink by the same amount.
AI cuts attack timelines from days to minutes, so detection-to-containment time has to shrink the same way
The rules for handling vulnerabilities are also starting to change. On Nov. 16, Anthropic announced it would shorten the deadline before it makes discovered vulnerabilities public. For reports made on or after Jan. 1, 2027, the deadline for critical and high-severity issues will drop from 90 days to 30. Anthropic said the reason is that frontier AI is good at finding flaws, and other AI systems quickly rediscover the same ones. The move cuts the industry-standard 90 days because of AI capabilities, and it points in the same direction as GTIG's report.
For Japanese companies, the realistic step is not to apply every patch within hours. First, list the software running on internet-facing servers and websites. For that software, set up a procedure to apply critical patches as the top priority. Also decide in advance on fallback protections, such as blocking network traffic, for the period before a patch can be applied.
What we still don't know
Some points about GTIG's report could not be confirmed. The report and coverage we reviewed (SiliconANGLE, TechTarget and others) did not say who was behind exploit_agent, how much damage it caused or which AI model it used. The comments from Elrod and Wyler come from press coverage and are not part of GTIG's report itself.
Sources
- https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold
- https://www.helpnetsecurity.com/2026/11/17/ai-powered-cyberattacks-gtig-report/
- https://siliconangle.com/2026/11/16/google-says-an-ai-framework-turned-security-patches-into-exploits-with-no-human-in-the-loop/
- https://www.techtarget.com/news/google-finds-ai-agents-autonomously-turning-patches-into-exploits/874837/