JupiteX Get the app
Defence & Security17 Nov 2026 · about 6 min

Google found hackers using autonomous AI agents to turn security patches into exploits

The brief

Google’s researchers found attackers using autonomous AI agents to move from a public security patch toward exploit code and attempts to run it. This matters because the workflow can automate analysis and testing that attackers have long done themselves. In one case, a publicly exposed staging server held a framework called exploit agent, stolen victim data, and scripts used after intrusions. It connected frontier and locally hosted language models. An orchestrator coordinated agents that compared patched and unpatched code, generated exploit candidates, and tried them against targets. The report does not identify which models handled each task or say how often the generated exploits worked. Google also could not determine whether exploit agent obtained the stolen data on the server. The framework’s capability still signals a faster route from patch information to attack attempts, though researchers said these systems generate detectable traffic and artifacts.

01

How do attackers use AI agents to turn a security patch into an exploit?

Google’s researchers found attackers using autonomous AI agents to move from a public security patch toward exploit code and attempts to run it. This matters because the workflow can automate analysis and testing that attackers have long done themselves.

In one case, a publicly exposed staging server held a framework called exploit_agent, stolen victim data, and scripts used after intrusions. It connected frontier and locally hosted language models. An orchestrator coordinated agents that compared patched and unpatched code, generated exploit candidates, and tried them against targets.

The report does not identify which models handled each task or say how often the generated exploits worked. Google also could not determine whether exploit_agent obtained the stolen data on the server. The framework’s capability still signals a faster route from patch information to attack attempts, though researchers said these systems generate detectable traffic and artifacts.

02

What is patch diffing, and how can comparing code before and after a patch reveal a vulnerability?

Patch diffing means comparing a program’s code before and after a security patch. The comparison can reveal which parts changed and offer clues about what problem the fix addresses. Attackers can use those clues to investigate whether the original weakness can be triggered.

For example, if a patch changes a section of code, examining that change may help identify the behavior the developers corrected. An attacker can then study the surrounding code and try to create an exploit candidate that takes advantage of the old behavior. The article does not identify the exact flaw or code changes in the exploit_agent case.

Attackers have used patch comparisons for years, so the technique itself is not new. What has changed is that AI agents can do much of the diffing and exploit-generation work themselves. Google says its observed framework also attempted to execute generated candidates, although researchers did not disclose how often they succeeded.

03

How quickly have AI-assisted attackers and researchers moved from a vulnerability disclosure to working exploits?

The article describes exploit development happening on a very short timeline after vulnerability information became public. After two PaperCut vulnerabilities were exploited as zero-days in August, security researchers and Huntress developed working exploits with LLM assistance within a day of disclosure.

A separate example came from GreyNoise. Four days after the initial disclosure, it documented a campaign in which an attacker likely used language models to develop exploits, build target lists, and attack hundreds of organizations. The attacker reportedly moved from initial access to domain administrator within minutes.

The article does not establish that every exploit in these cases was generated by AI, and it says many purported PHP exploits posted after patches shipped did not work. Still, the examples show that working exploits and attack attempts can follow disclosures quickly. Defenders may have little time to assess newly public information and deploy fixes.

04

What does faster exploit development mean for organizations trying to install security fixes?

Faster exploit development shortens the time organizations have to install security fixes before attackers try to use the underlying weakness. Google describes this as a concern for application security teams: attackers can move quickly from a published patch to a working exploit.

The article gives several examples of compressed timelines. Researchers and Huntress developed working PaperCut exploits with LLM assistance within a day of disclosure. GreyNoise documented a likely AI-assisted campaign four days after an initial disclosure; the attacker reached domain administrator within minutes of initial access. These examples show that the exposure can escalate rapidly once an attack begins.

Organizations therefore face pressure to identify relevant fixes and assess exploit claims quickly. But speed is not the only problem: hundreds of purported exploits for two PHP vulnerabilities appeared after patches shipped, and many did not work. Security teams had to distinguish real threats from nonfunctional code, while attackers’ AI systems could also produce detectable traffic and artifacts.

05

How do an orchestrator and specialized agents divide up work in an automated attack framework?

An orchestrator is the coordinating part of an automated framework. It assigns or coordinates tasks, shares context among agents, and invokes tools through interfaces such as MCP. Specialized agents can then focus on parts of an offensive security workflow rather than one agent doing everything.

Google describes exploit_agent as using multiple models to automate patch diffing, generate exploit candidates, and attempt to execute them against targets. Other frameworks divide work differently. HexStrike AI, for example, exposes more than 150 offensive security tools through MCP, with specialized agents for reconnaissance, exploitation, and binary analysis. CAIRN coordinates agents through a shared blackboard and an observe-orient-decide-act loop.

These designs let a framework use results from one task to guide what happens next. Google observed CAIRN in intrusions targeting U.S. aerospace organizations. The article says these frameworks can generate substantial network traffic and recognizable artifacts, and agents may hallucinate technical details or nonexistent targets. Their current trade-off is velocity over stealth.

06

How is using AI to analyze patches different from the patch-comparison techniques attackers have used for years?

The basic technique is not new: attackers have long compared code before and after a security patch to work out what was fixed and how the original weakness might be exploited. This process, called patch diffing, uses the changes in the code as clues.

The newer step is automation. In Google’s observed exploit_agent framework, language models connected through a multi-agent system helped automate patch diffing, produce exploit candidates, and attempt to execute them. An orchestrator coordinated specialized agents, shared context, and invoked tools. The agents applied familiar coding workflows to offensive security tasks.

This shifts how much work can be done without direct human approval, but it does not mean every generated exploit works. Google did not report success rates or identify which models performed individual tasks. Researchers also found that these systems can hallucinate details and targets, and they leave detectable traffic and code artifacts. The change is automation and speed, not the invention of patch comparison.

07

What is a software vulnerability, and how does a security patch fix it?

In general software-security terms, a vulnerability is a weakness in software that can allow it to be used in an unintended or unsafe way. The article does not provide a formal definition, but it describes attackers looking for vulnerabilities in software and trying to exploit them.

A security patch is a change intended to fix a software problem. Comparing code before and after a patch can show which parts changed and give clues about the weakness the developers addressed. Attackers can study those clues to make exploit candidates, while defenders apply the patch to reduce the risk from the flaw.

The article’s central concern is that agents can automate parts of this process: finding clues in patches, generating exploit code, and attempting to run it. Google did not disclose how often generated exploits succeeded. It also found that agents sometimes hallucinate technical details or nonexistent targets, so automated output is not necessarily a working exploit.

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