AI Exploits Are Being Built Faster From Security Patches
AI exploits are becoming faster to build and deploy as threat actors move beyond asking artificial intelligence for individual answers. Google Threat Intelligence Group (GTIG) says attackers are using autonomous, multi-agent frameworks, meaning several software agents that can perform connected tasks with limited human direction, to carry out multi-stage intrusions at greater speed.
The frameworks named by GTIG include ARTEX, CAIRN, CyberStrikeAI and PentAGI. Threat actors are also abusing open-source penetration-testing frameworks, which are tools designed to assess security, to automate individual stages and operational tasks during live attacks. Google says financially motivated groups and suspected espionage actors aligned with China have been seen experimenting with these frameworks and related systems.
According to GTIG, the wider change is that attacks can now be organised at machine speed. The underlying attack methods may not be new, but automated tools can connect different attack paths and scale operations more quickly. Open-source frameworks also reduce the technical barrier for less-skilled attackers, making it easier to integrate exploits, expand attacks and create convincing phishing campaigns.
A major area of concern is the use of large language models, or LLMs, which are AI systems trained to generate and analyse language and code. Attackers are using LLMs and agentic frameworks, systems that allow AI tools to plan and perform tasks, to reverse engineer software patches through differential analysis. This means comparing a patched version with an earlier version to identify what changed and infer the vulnerability that was fixed.
That process can help attackers weaponise newly disclosed flaws quickly. Google observed this activity in connection with vulnerabilities in PaperCut NG/MF earlier this September. GTIG says attackers may find it more efficient to focus an LLM on a limited collection of material, including different source-code versions, pre-release or preview versions, release notes, security research and proof-of-concept code. A proof of concept, or POC, is a working demonstration that shows how a vulnerability can be exploited.
This can compress the period between a patch being published and active exploitation. The affected software may already have a fix available, but organisations still need to apply it before attackers can use information revealed by the change. For website owners and IT teams, the finding makes the speed of patch management particularly important, because a public fix can also provide useful clues about the flaw it addresses.
Google also described misuse of other services and frameworks. ZeroHollow is an inference service used for exploit development and for routing requests between LLMs. Agent Zero is an open-source agent framework that includes autonomous code execution and dynamic tool generation, allowing it to create or select tools as tasks change.
Attackers are targeting AI infrastructure for financial gain as well. In LLMjacking attacks, they steal LLM credentials, which are the access details used to connect to AI services, and cloud resources from compromised enterprise environments. The stolen access can then be resold through underground proxy services such as APIMart and JetNight at reduced prices. These resale services are also known as LLM mixers.
Phishing is being automated in parallel. Jieyou, a Chinese-language phishing platform, has added an LLM chatbot assistant that lets customers configure phishing panels and alter existing templates using natural-language commands. Phishing is the use of deceptive messages or pages to trick people into giving up information or access.
GTIG says this reflects a broader trend among phishing-as-a-service, or PhaaS, providers. These services supply phishing infrastructure to customers who may not have the skills to build it themselves. Instead of editing payloads manually, lower-skilled affiliates can use chatbot instructions to customise deceptive stories and phishing interfaces while an operation is under way. The result is a shorter path from an attacker’s intention to a tailored phishing campaign.
Businesses can reduce exposure by applying security patches promptly, reviewing which employees and services can access AI and cloud accounts, and protecting account sign-ins with multi-factor authentication, which requires an additional confirmation beyond a password. Website owners should also keep their hosting software and connected applications up to date, while staff should treat unexpected login requests and urgent messages with caution. AEU-I provides security-first IT, infrastructure and consulting, which is relevant for organisations reviewing these risks across their systems.
The findings do not mean that every attack is fully autonomous, but they show how AI agents and open-source tools can automate enough separate tasks to make exploitation and phishing faster, broader and more accessible to attackers.