BitCorner

We question legacy finance and explore Bitcoin as the alternative.

News › Security

Security

Google says an AI framework turned security patches into working exploits with no human in the loop

Google's Threat Intelligence Group documented 'exploit_agent', which compares code before and after a patch, drafts an exploit and runs it automatically.

by Juan Cienfuegos·
BitCorner

Google Threat Intelligence Group (GTIG) says it has found the first case of attackers using an AI framework to turn security patches into working exploits with no human involved. In its report, titled "Agentic Threat Activity Crosses a Significant Threshold," it calls the tool exploit_agent, according to SiliconANGLE.

How it works: exploit_agent continuously compares the code of popular web applications before and after each patch announcement. When it finds a likely flaw, it drafts an exploit and runs it automatically. It ran on an exposed staging server and was connected to several frontier and locally hosted large language models.

The same server held the stolen data from victims and scripts for brute-forcing, installing webshells and removing webshells left by rival attackers. Victim telemetry suggests the operator focused on French websites in media, entertainment and education. A configuration-file comment in Russian told the user to enter keys "right here, not in the chat," but the article does not state a firm attribution.

GTIG's reading is that two months ago it had not seen fully autonomous attack pipelines used against real targets, and that changed in the latest quarter. Because patched flaws are a finite source, it says attackers may find it more efficient to target n-day vulnerabilities, flaws that are already public, than to hunt for zero-days. It also notes that unrestricted autonomous frameworks leave a noisy forensic trail, which makes them easier to detect.

The report lists related cases: two chained PaperCut flaws exploited as zero-days in late August, with researchers and Huntress Labs using LLMs to generate working exploits within a day of disclosure; a GreyNoise-observed campaign that built exploits four days after disclosure and hit hundreds of organizations; a multi-agent framework called CAIRN used against the US aerospace sector; and a suspected North Korean group, UNC6836, that altered `.claude/settings.json` files in internal repositories so each launch of Anthropic's Claude Code downloaded a backdoor.

Google says it disabled the accounts and assets tied to the Gemini misuse it found, and that Google DeepMind used the findings to strengthen its classifiers. For defenders, the lesson is about time: if an exploit can appear within hours of a patch, the window to install updates is shorter than the monthly cycle many teams still run on.

Facts only, never financial advice.

About the author

Juan Cienfuegos

Founder & CEO, BitCorner Media

Host of BitCorner, the Bitcoin podcast from El Salvador. BitCorner Media is the first Bitcoin and tech media company of El Salvador and helps Bitcoin and tech companies reach and educate Spanish-speaking audiences.