JupiteX Get the app
Defence & Security18 Nov 2026 · about 7 min

Google Finds Hackers Using AI Agents to Build and Run Exploits From Software Patches

The brief

AI systems are helping attackers turn software fixes into attack instructions. Google says threat actors use large language models and agentic frameworks to automate patch reverse engineering, then rapidly weaponize newly disclosed vulnerabilities. That can shrink the time defenders have to respond. The key technique is differential analysis: compare software versions from before and after a fix and look for what changed. Attackers can use those differences, along with release notes, security research, and proof-of-concept code, to infer the flaw and develop an exploit. Google says this can be efficient because researchers can focus on a finite set of materials. PaperCut NG/MF flaws were an example of this activity earlier this September. The wider concern is that automated pipelines can compress the gap between patch disclosure and active exploitation. Organizations may face attacks before completing updates. Google does not say that every patch is targeted or that these tools independently carry out an entire intrusion.

01

What are hackers using AI agents to do with software patches?

AI systems are helping attackers turn software fixes into attack instructions. Google says threat actors use large language models and agentic frameworks to automate patch reverse engineering, then rapidly weaponize newly disclosed vulnerabilities. That can shrink the time defenders have to respond.

The key technique is differential analysis: compare software versions from before and after a fix and look for what changed. Attackers can use those differences, along with release notes, security research, and proof-of-concept code, to infer the flaw and develop an exploit. Google says this can be efficient because researchers can focus on a finite set of materials.

PaperCut NG/MF flaws were an example of this activity earlier this September. The wider concern is that automated pipelines can compress the gap between patch disclosure and active exploitation. Organizations may face attacks before completing updates. Google does not say that every patch is targeted or that these tools independently carry out an entire intrusion.

02

What is an autonomous, multi-agent offensive framework?

An autonomous, multi-agent offensive framework is a system that uses AI agents or automated components to carry out offensive cyber tasks. In this context, “autonomous” means tasks can be handled with less hands-on direction, while “multi-agent” points to multiple agents or components working within a framework. The article describes these tools as a move beyond manual prompting.

Google names ARTEX, CAIRN, CyberStrikeAI, and PentAGI as examples. It says threat actors use open-source agentic penetration-testing frameworks to automate discrete phases and specific tasks during live intrusions. The article does not detail the internal design of these named tools, but it describes frameworks being used to support multi-stage operations.

The significance is that automation can let attackers operate at greater speed and scale. They can chain varied attack paths while the underlying mechanics of attacks remain similar. Google has observed experimentation by financially motivated actors and suspected espionage actors aligned with China. These frameworks lower the barrier to entry, but the source does not claim they autonomously perform every step of an attack.

03

How can comparing software versions before and after a patch help attackers build an exploit?

When a software maker releases a patch, comparing the affected version with the fixed version can reveal which code changed. If attackers identify a change that closes a weakness, they can use it to understand where the flaw was and how it might be triggered. This process is called differential analysis in the article.

Google says threat actors can combine version comparisons with release notes, security research, and proof-of-concept code. They may also examine source code from pre-release or preview channels. LLM-driven pipelines can help automate this labor-intensive analysis, making it easier to move from a patch to a working exploit. The article cites PaperCut NG/MF flaws as an example of rapid weaponization activity.

This matters because the fix itself can provide useful information to attackers, even as it protects systems once installed. Google says this approach can compress the window between patch disclosure and active exploitation. It does not provide technical details of the PaperCut exploit or claim that every version comparison leads to a successful attack.

04

How can AI change the speed and scale of cyberattacks?

AI can change cyberattacks by automating individual tasks that once required substantial manual effort. Google describes a shift toward attacks taking place at machine speed. Agentic frameworks and large language models can support tasks such as vulnerability analysis, patch reverse engineering, and proof-of-concept weaponization.

For example, attackers can compare software versions around a patch, inspect related research, and use AI-driven pipelines to develop an exploit. Open-source frameworks can also automate selected phases during a live intrusion. Google says operations can be scaled by chaining together varied attack paths, even when the underlying mechanics of the attacks remain the same.

The result can be faster operations and a lower barrier to entry. The article also describes LLM-assisted phishing, where customers can use natural language to change templates and configure phishing panels. It does not claim AI replaces human attackers or makes every attack fully autonomous. Instead, these tools can speed up particular tasks and help more actors carry them out at greater operational velocity.

05

What can happen when attackers weaponize a vulnerability before organizations install its patch?

If attackers build an exploit before an organization installs a patch, systems that still contain the flaw may remain exposed. In general, a successful exploit can let an attacker carry out unauthorized actions on a vulnerable system. The specific damage depends on the flaw and the target; the article does not describe resulting harm to PaperCut customers.

Google says AI-driven pipelines can help attackers analyze patched vulnerabilities by comparing source code versions and reviewing release notes, security research, and proof-of-concept code. That can help them weaponize a vulnerability after its fix is disclosed. The article points to PaperCut NG/MF flaws as an example of this kind of activity.

The concern is a shorter period between public disclosure of a patch and active exploitation. Organizations that have not installed the fix may have less time to respond. The article does not say how many organizations were affected or identify particular victims. Its broader warning is that automated analysis can make the transition from a disclosed flaw to an attack faster.

06

How do criminals make money by stealing access to AI services and reselling it?

LLMjacking is the abuse of access to large language model services. In the activity described by Google, attackers steal LLM credentials and cloud resources from compromised enterprise environments. They can then sell the access to others, creating a way to profit from resources they did not legitimately obtain.

The article names APIMart and JetNight as underground proxy services used to resell this access at discounted prices. These services are also known as LLM mixers. The source does not describe the exact payment process, the prices charged, or how the stolen credentials are obtained in each case; it says the access and resources come from compromised enterprise environments.

This means AI infrastructure itself is a target, not just a tool attackers might use. Stolen access can be offered to customers through proxy services, extending its use beyond the original compromise. Google presents LLMjacking alongside other forms of AI-related misuse, including automated exploit development and phishing. The article gives no figures for the scale or profits of these resale operations.

07

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

A software vulnerability is a weakness or flaw in a program that could allow unintended or unauthorized behavior. A patch is a software update designed to correct a problem, including a security vulnerability. These are general definitions; the article focuses on attackers analyzing patches rather than defining either term.

When a patch changes code to fix a vulnerability, comparing the earlier and later versions may show what was altered. Google says attackers can use that comparison, together with release notes, security research, and proof-of-concept code, to work out how a flaw might be exploited. The article calls this approach differential analysis and describes AI-assisted automation of the process.

Installing a relevant patch can address the weakness in the affected software. However, Google warns that threat actors may rapidly weaponize patched vulnerabilities, compressing the time between disclosure and active exploitation. This makes timely attention to updates important. The article does not explain how organizations should prioritize or deploy patches, nor does it give details about the specific code changes in the PaperCut NG/MF example.

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