Google: Attackers Now Use AI Agents to Turn Software Patches Into Working Exploits
Google says criminals' use of artificial intelligence has moved beyond typing prompts into a chatbot. According to the Google Threat Intelligence Group (GTIG), attackers are now deploying autonomous, multi-agent frameworks such as ARTEX, CAIRN, CyberStrikeAI and PentAGI to carry out multi-stage intrusions at what Google calls machine speed. Financially motivated actors and suspected espionage actors aligned with China have been seen experimenting with these tools.
A second trend concerns patches. Attackers are using large language models (LLMs) and agent frameworks to compare software versions, a process known as differential analysis, and work out what a security fix changed. That lets them build exploits for newly disclosed flaws quickly. Google pointed to the PaperCut NG/MF vulnerabilities in September as an example. It notes that attackers can focus on a limited set of material, including source code versions, release notes, security research and proof-of-concept code, which shrinks the gap between a patch being released and active exploitation.
Google also describes attackers targeting AI infrastructure itself. In so-called LLMjacking, criminals steal LLM credentials and cloud resources from compromised enterprise environments, then resell access at a discount through underground proxy services such as APIMart and JetNight, known as LLM mixers.
AI is also lowering the skill needed for phishing. A Chinese-language phishing platform called Jieyou has added an LLM chatbot that lets customers configure phishing panels and edit templates using plain language commands. Google says low-skilled affiliates can now customise lures and fake login pages on the fly, which it sees as part of a wider trend of phishing-as-a-service providers using LLMs as a force multiplier.
Why it matters
The assumption that businesses have days or weeks after a patch release before exploitation begins is weakened when AI can analyse the patch and build an exploit rapidly. Organisations running internet-facing software such as PaperCut are most exposed, and patch timelines may need to shrink accordingly. Stolen cloud and LLM credentials now have resale value, so any cloud account with AI services attached is a target in its own right.
What to do
- Prioritise patching of internet-facing software such as PaperCut NG/MF as soon as fixes are released.
- Review cloud accounts for unexpected LLM or AI service usage that could indicate LLMjacking.
- Rotate and restrict LLM API keys and cloud credentials, and monitor them for unusual billing or activity.
- Train staff to expect polished, convincing phishing lures, since AI makes customised templates easy to produce.