Google says hackers use AI agents to find and exploit software flaws
Google’s Threat Intelligence Group said it had identified attackers using large language model frameworks to automate vulnerability discovery and exploitation without a person directing each step. The systems can examine software, identify possible flaws, and generate and run code designed to exploit them. That matters because automation can speed up the path from finding a weakness to attempting an attack. GTIG described a multi-agent framework on an exposed staging server. It compared software code before and after patch announcements for popular web applications. The system used those differences to look for vulnerabilities, then could autonomously develop and execute exploit code. The report described this as a continuous process, rather than a person carrying out each task manually. Google’s finding was a first for GTIG, but it did not say that fully autonomous pipelines had already been observed attacking targets in the wild. Its September report had said it had not yet seen that. The newer findings suggest defenders may have less time to apply patches before attackers can develop ways to exploit flaws.
What did Google say the attackers’ AI systems could do without step-by-step human direction?
Google’s Threat Intelligence Group said it had identified attackers using large language model frameworks to automate vulnerability discovery and exploitation without a person directing each step. The systems can examine software, identify possible flaws, and generate and run code designed to exploit them. That matters because automation can speed up the path from finding a weakness to attempting an attack.
GTIG described a multi-agent framework on an exposed staging server. It compared software code before and after patch announcements for popular web applications. The system used those differences to look for vulnerabilities, then could autonomously develop and execute exploit code. The report described this as a continuous process, rather than a person carrying out each task manually.
Google’s finding was a first for GTIG, but it did not say that fully autonomous pipelines had already been observed attacking targets in the wild. Its September report had said it had not yet seen that. The newer findings suggest defenders may have less time to apply patches before attackers can develop ways to exploit flaws.
How can an AI agent use changes between software versions to find a newly fixed flaw and try to exploit it?
A software patch changes code to address a weakness. Comparing versions from before and after that change—a method called differential analysis—can point to the part of the software that was fixed. Attackers can use that clue to investigate the underlying vulnerability and try to build an exploit. The article says AI agents can automate this work, making it faster.
GTIG found a framework set up to continuously compare code before and after patch announcements for popular web applications. It identified potential flaws and could generate and execute exploit code. This is especially relevant to N-day vulnerabilities, which the article describes as publicly known security flaws: patch information can help attackers study what was fixed and attempt to exploit systems that remain vulnerable.
The framework was found on an exposed staging server, and Google described it as capable of autonomous discovery and exploitation. The report warns that this process can leave organizations less time to install patches. It does not say that the specific framework was observed attacking targets in the wild, so its described capability should not be mistaken for proof of such deployment.
What is an N-day vulnerability, and how is it different from a zero-day vulnerability?
The article calls N-day vulnerabilities “publicly known security flaws.” The “N” indicates that some time has passed since the flaw became known or was disclosed. Once details are public, attackers may study the flaw and target systems that have not yet been patched. That is why N-days can remain dangerous even after a fix is released.
In standard cybersecurity usage, a zero-day vulnerability is a flaw being exploited before a patch is available to protect against it. “Zero-day” signals that defenders have had no days with a fix available before the exploit. The terms describe different situations around knowledge and patch availability, not different kinds of software.
The article discusses both categories in Google’s figures for exploited vulnerabilities. It reports an average of eight zero-days exploited per month in 2025 and 11 per month from January through August 2026. It also describes attackers using patch comparisons to study N-day flaws. The article does not give a separate technical definition of zero-day; that explanation reflects standard cybersecurity usage.
How many zero-day vulnerabilities were exploited per month on average in 2025 and from January through August 2026, and what was the August count?
Google’s figures show a rise in exploited zero-day vulnerabilities across the periods it reported. The average was eight per month in 2025. From January through August 2026, it averaged 11 per month. The August count reached 22. These numbers matter because zero-days are flaws being exploited before a fix is available, so a spike can signal intense pressure on defenders.
The report also described a broader increase in vulnerability disclosures: they more than doubled between January and August, according to Google. The article does not say that all disclosed vulnerabilities were zero-days, or that every exploited flaw was used in an attack against the same type of system. It distinguishes zero-days from N-day flaws, which it describes as publicly known security flaws.
The figures cover the periods stated in the report, rather than forecasting what will happen next. Google also warned that AI-assisted workflows can help attackers discover and exploit flaws more quickly. That speed may reduce the time organizations have to apply patches, making quick attention to security updates especially important.
Why do automated discovery and exploitation give organizations less time to patch vulnerable systems?
A patch is a software update that addresses a flaw, but organizations need time to apply it to affected systems. Google warns that attackers’ automated workflows can compress the time between a patch announcement and attempts to exploit the vulnerability. The less time defenders have to update their systems, the longer vulnerable systems may remain exposed.
GTIG described a framework that compared software code before and after patch announcements. By examining those differences, it could identify potential vulnerabilities and generate and execute exploit code. Google said attackers use AI agent frameworks to automate patch reverse engineering and weaponize N-day vulnerabilities, which are publicly known flaws. The report warns that such pipelines can target internet-facing servers without human intervention.
The report’s concern is not only that attackers can work faster, but that organizations may have less time to respond. Google said vulnerability disclosures more than doubled between January and August. It also reported rising zero-day exploitation figures. These trends add urgency to patching, although the article does not prescribe a particular patching schedule or describe how quickly every organization responds.
What can happen when attackers exploit a flaw in an internet-facing server and move from initial access to administrator control?
An internet-facing server is reachable from the internet, and the article warns that attackers can scale exploitation pipelines against such systems without human intervention. If attackers exploit a vulnerability to gain initial access, they may then try to expand their control. Google’s report says automated workflows can accelerate the move from that first access to domain administrator access.
The article gives an August 2026 PaperCut campaign as an example. Cybersecurity firm GreyNoise said the attackers achieved domain administrator access within minutes. This illustrates how quickly an intrusion can escalate after an attacker gets in. The article does not specify the exact steps used in that campaign, so it does not support a more detailed account of how the attackers gained control.
Domain administrator access represents a high level of control in the affected domain. The report’s broader concern is that AI-assisted discovery and exploitation can speed up attacks, leaving defenders less time to respond. Paradox separately warned that agentic AI operations can automate much of the cyberattack cycle, from discovery to exploitation and data theft. Those findings point to a fast-changing threat, not a guarantee that every exploited server leads to administrator access.
What is a software vulnerability, and how does installing a patch help protect a system?
A software vulnerability is a weakness or flaw in a program that can create a security risk. The article describes attackers identifying such flaws in software and attempting to exploit them. It does not provide a general definition, so this explanation uses established cybersecurity terminology. A vulnerability may affect popular applications or systems exposed to the internet, as in the examples discussed by Google.
A patch is a software update intended to correct a vulnerability or otherwise address a security issue. Installing it brings the affected software to a version with that correction, helping protect the system from attacks that rely on the flaw. The article describes attackers comparing code before and after patch announcements, then trying to exploit systems that have not been updated.
Google’s warning is that automated discovery and exploit development can leave organizations less time to apply patches. The report says attackers can use AI frameworks to analyze software changes and generate exploit code. Installing relevant patches can close the weakness those exploits target, but the article does not explain particular patching procedures or guarantee that a patch prevents every kind of attack.
This brief was written by AI from the original reporting and checked by other models. Names, figures and quotes come from the source; read it for full context.
Read more in the JupiteX app
Pulse is free. New stories every 4 hours, each one broken into the questions that explain it.
Or read more news on the web