Advertisement

AI

Google says an AI agent turned security patches into working exploits with no human involved

Share this story

Google's threat researchers found an attack framework that reads software patches, writes exploits and launches them on its own. Here is what changed.

By Emily4 min read

A hand typing on a laptop in a dark room, the screen filled with green code, washed in a green-to-violet tint, with the headline AI agent turned patches into exploits, no human in the loop.

For years, security people have warned that AI would eventually let attackers go from a published software fix to a working attack without anyone sitting at the keyboard. Google now says it has seen exactly that. In its latest quarterly AI threat report, published on Monday 16 November, the Google Threat Intelligence Group (GTIG) says that in the past quarter it "identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time".

The report is titled, without much understatement, "Agentic Threat Activity Crosses a Significant Threshold".

What Google found

The tool at the centre of it is called exploit_agent. GTIG says it found the framework on an exposed staging server, where it was wired into "multiple frontier and local LLMs", in other words both commercial AI models and ones running on the attacker's own hardware.

According to Google, the framework was set up to run continuous comparisons of "popular web application software". It looks at the code before and after a patch is announced, works out what the fix was hiding, drafts an exploit and then runs it, all without a person stepping in. Google did not name the software involved.

It was not just a lab experiment. The same server held data taken from victim organisations, along with scripts for internal reconnaissance, brute-forcing passwords, installing webshells and, in a detail that says a lot about the neighbourhood, removing webshells left behind by other attackers. GTIG says victim telemetry on the server suggests the operator was particularly interested in French websites in media and entertainment, and in the education sector.

Who was behind it is less clear. A configuration file carried Russian-language comments, one of which, in Google's translation, tells the operator to enter their AI API keys "right here, not in the chat". That points to a suspected Russian speaker, but Google has not named a group or a country.

Advertisement

Why patches are the target

The interesting part is the strategy. Google argues that attackers may find it more efficient to point AI at "n-day" flaws, ones that have already been fixed, than to hunt for brand-new zero-days. A patch gives the model a small, finite pile of material to study: the old code, the new code, the release notes and any public proof-of-concept. That shrinks the gap between a fix being published and attacks starting.

The report gives a recent example. In late August, attackers exploited two flaws in PaperCut print management servers that, when chained together, allowed remote code execution without logging in. Google says security researchers and Huntress used AI to build working exploits within a day of disclosure, and that GreyNoise reported a campaign four days after disclosure in which an attacker likely used AI to build exploits and hit hundreds of organisations.

GTIG also says the wider move to automation is well under way. It lists open-source attack frameworks, including ARTEX, CAIRN, CyberStrikeAI, HexStrike AI and PentAGI, that it has seen criminals or suspected state-linked groups use or experiment with. Google's view is that these tools currently "trade stealth for velocity": they are fast but noisy, which makes them easier to spot. It expects that to change as the models improve.

Other findings in the report

The report covers more than exploits. Mandiant, Google's incident response arm, says it has dealt with several "LLMJacking" cases this year, in which criminals steal cloud credentials to run up AI usage on someone else's account. Google says single-day fraudulent charges exceeded $100,000 in some cases, and in the worst cases reached $1 million before the abuse was stopped. That access is then resold through what GTIG calls "LLM Mixers", underground shops selling cheap access to AI models.

What it means for UK readers

Nothing here is a reason to panic, and Google does not claim the framework hit UK targets. The practical lesson is an old one, made more urgent: the time between a patch appearing and someone trying to exploit it is getting shorter. If you run a website, a small business network or a self-hosted app, applying updates quickly matters more than it did a year ago. For everyone else, it is a good moment to switch on automatic updates for your phone, laptop and router, and to treat any unpatched, internet-facing kit as the first thing an automated attacker will find.

This also follows yesterday's news that California is suing OpenAI over rogue AI agents. That case is about AI labs losing control of their own models. This one is about criminals deliberately building agents to attack. Both point the same way: AI agents are now doing real work on the internet, and not all of it is welcome.

Sources

Cover photo: Sora Shimazaki / Pexels.

Share this story

Spot an error? Tell the desk via the contact page. How we score devices is in the editorial policy.

The status, occasionally

A short UK briefing when a phone, tablet or wearable is actually worth your money. No daily spam.

Occasional briefings. Unsubscribe any time. We only store your address if newsletter storage is connected.

Advertisement

Cookies on TechnoStatus

We use a necessary cookie to remember this choice. Analytics and advertising cookies stay off unless you opt in. Rejecting is as easy as accepting. Read the cookie policy and privacy policy.