Itdaily - Google sees AI agent turn patches into exploits without human intervention

Google sees AI agent turn patches into exploits without human intervention

Hands of a man typing on a laptop, with colorful code on the screen.
Image: Tero Vesalainen/Shutterstock

Attackers have for the first time deployed an AI framework that turns security patches into working exploits without human involvement, reports the Google Threat Intelligence Group. The framework is called exploit_agent and ran on an exposed server.

The Google Threat Intelligence Group (GTIG) published its new quarterly report on AI misuse on Monday, in collaboration with Mandiant. According to the report, a fully autonomous chain from patch analysis to execution of the attack has now also been observed outside the lab.

Two months ago, in the previous edition, GTIG still wrote that it had not seen fully autonomous attack chains against real targets. That changed in the past quarter.

How exploit_agent works

GTIG found exploit_agent on an exposed staging server. The framework is linked to several frontier models and locally hosted language models. It compares the code of popular web applications before and after each patch, looks for a likely vulnerability, writes an exploit and runs it automatically.

The same server held victim data, along with scripts for brute-force attacks, installing web shells and removing web shells planted by rival attackers. The telemetry suggests the operator mainly targeted French websites in media, entertainment and education. A Russian comment in the configuration file asks the user to enter the API keys “here, not in the chat”.

According to GTIG, a patched flaw is more efficient for attackers than a zero day. With a patch, there is a limited amount of source material: old and new code, release notes, research and proofs of concept. Anyone hunting for zero days has to cast a much wider net.

From patch to attack in a few days

The report illustrates the speed with two vulnerabilities in PaperCut print management servers. Attackers exploited them as zero days in late August, in combination with each other. Researchers and the security company Huntress used language models to build working exploits within a day.

GreyNoise saw a campaign against hundreds of organisations four days after disclosure, in which the attacker probably used language models to build and test exploits. A single entry point grew into domain admin within minutes.

For two PHP vulnerabilities (CVE-2026-109618 and CVE-2026-109619), too, a proof of concept and patched code appeared before the official patch of 21 October. Hundreds of purported exploits then surfaced that did not work and were probably AI-generated. Defenders had to check them first.

Between January and August 2026, the number of disclosed vulnerabilities and exploits nearly doubled. Zero-day exploits rose from an average of eight per month in 2025 to eleven in 2026, with a peak of 22 in August, writes The Register.

Other frameworks and AI misuse

GTIG also sees open source penetration testing frameworks in real attacks. CAIRN, a multi-agent framework, turned up in a breach in the US aerospace sector. Between late August and early October, an attacker used a customised platform based on PentAGI to scan organisations in Asia for vulnerabilities. CyberStrikeAI and HexStrike AI are also being tested. These tools leave traces in network traffic and code, which makes them easier to detect.

With LLMjacking, attackers steal credentials for language models and cloud resources and resell the access, writes Security Affairs. Mandiant saw fraudulent charges of more than 100,000 dollars in a single day, and in the worst case 1 million dollars before the abuse was stopped, reports The Next Web.

In influence operations, a commercial contractor uses the Gemini API to build a platform for more than 100,000 social media accounts. A suspected North Korean group, UNC6836 according to the source material, also modified .claude/settings.json files in internal code repositories, so that Claude Code downloaded the SOMBERMEME backdoor on every launch.

Google’s countermeasures

Google says it has disabled the accounts and assets behind the misuse of Gemini and that Google DeepMind is using the findings to strengthen its classifiers. Google SecOps has released detection rules for ARTEX and Mandiant published a hardening guide. GTIG expects attackers to keep deploying these frameworks, using fewer people and more cunning as the tools improve.