Google's new AI threat report explained: What are these autonomous AI cyber attacks

Google Threat Intelligence Group (GTIG) published their AI threat tracker report today which was majorly focused on the use of autonomous LLM and agentic frameworks to scan updates and patches coming to softwares, figure out what flaws they were for and then build an attack specifically for that without any human involvement. Google found that just two months prior to this, there weren't any fully autonomous attack pipelines like this targeting the wild digital jungle of AI.

Meet exploit_agent

The most obvious example would be exploit_agent, which is deployed on an exposed staging server connected to a variety of language models on the frontiers and locally. This framework analyses popular web application source code both prior and after the announcement of each fix, and once a possible vulnerability has been detected, the exploit generated by this system operates automatically. The very same server contained stolen information of victims and tools for deployment of webshells. According to the victim telemetry, the operator was particularly interested in French media and educational web applications.

Why patches are the new target

Generally, security patches will address an issue, and by addressing the issue, give away exactly where it was found. Hackers will be able to use that knowledge and target everyone who has not yet applied the patch. Known vulnerabilities are referred to as n-day issues and are in contrast to zero-day issues, where the vendor does not yet know the vulnerability exists. With a limited source code for any patched issue, GTIG hypothesizes that hackers would prefer pointing LLMs at n-day vulnerabilities rather than zero-day vulnerabilities.

There are already signs of that happening. Back in late August, two chain vulnerabilities were exploited on PaperCut print management servers. Language models were used to create working exploits just one day after the disclosure of the issue, and a campaign that targeted organizations four days after disclosure was likely using LLM created exploits.

It isn't just one tool

Open-source frameworks are also being used to automate steps of live intrusions. One actor used the multi-agent framework CAIRN against the US aerospace sector, while another ran a customised PentAGI platform to scan organisations across Asia for vulnerabilities. The report also points to attacks aimed at developer tools: a suspected North Korean group altered .claude/settings.json files in internal repositories so that launching Claude Code would download its backdoor.

The report isn't only about hacking. One commercial contractor was building a platform, with help from the Gemini API, designed to run more than 100,000 social media accounts. Fake accounts, convincing photos and automated posting at that scale make coordinated campaigns harder to spot.

What it means for you

The findings aren't specific to any one country, but the advice is practical. Turn on automatic updates for phones, laptops, routers and smart devices, and apply security patches as soon as they arrive. Businesses running internet-facing software, particularly older or unmanaged systems, have the most to lose from slow patching. Developers should treat settings files and configuration in code repositories as part of their attack surface.

A note of caution

There is some reassurance. The noisy forensic trail left by unrestricted autonomous execution is said to make these frameworks easier to detect. Google also says it disabled the accounts and assets tied to the Gemini misuse it found, and DeepMind used the findings to strengthen its classifiers.

It is also important to note who is reporting this. Google did announce the findings with its own product pitch, Google AI Threat Defense, which combines Gemini and other frontier models, Wiz, CodeMender and Mandiant. The evidence comes from Google's own telemetry, and how widespread this activity is beyond what Google could see remains unclear.

Vyom Ramani

Vyom Ramani A journalist with a soft spot for tech, games, and things that go beep. While waiting for a delayed metro or rebooting his brain, you’ll find him solving Rubik’s Cubes, bingeing F1, or hunting for the next great snack.

View full profile