Friday, 20 November 2026
2 agent hacks today▼ 9 vs yesterday (11)

Google finds AI agent building working exploits from patches

Threat hunters discovered an exposed staging server running a multi-agent framework that compared code before and after updates, then attacked on its own.

By The Agentic Times ·

Google finds AI agent building working exploits from patches
· Image: theregister.com

Google's Threat Intelligence Group (GTIG) said on 16 November 2026 that it found an AI agent that turned security patches into working exploits with no human in the loop, and that the operation went on to steal data from victim organisations. Google described it as a first such occurrence. The findings were published in the company's fourth-quarter 2026 AI threat tracker, which said the case marks a significant advance in fully autonomous vulnerability discovery and exploitation compared with the previous quarter.

According to the report, GTIG's hunters found an exposed staging server hosting a multi-agent vulnerability research and weaponisation framework that plugs into multiple frontier and local models. The framework, called exploit_agent, used Russian-language tooling, and Google attributes the activity to a suspected Russian-speaking operator. It was configured to run continuous differential analysis of "popular web application software" — that is, comparing code before and after a security update to work out what the patch fixed. From there, the report said, it identified potential vulnerabilities, developed exploits, and executed them automatically without any human involvement.

The same server held data stolen from victims, Google said, along with "multiple" post-compromise scripts. Those scripts supported internal reconnaissance, brute-forcing, installing webshells (small scripts that give remote control of a web server), and removing webshells left behind by other attackers. "Victim telemetry on the server suggests the operator was specifically interested in targeting French websites associated with the media and entertainment industries, as well as the education sector," according to the report.

The tracker places the find in a broader pattern. Google said adversarial use of AI has moved beyond people typing prompts into a chatbot during an attack, towards autonomous multi-agent offensive frameworks used to run multi-stage intrusions at greater operational velocity. The company named several such frameworks in circulation, including the open source penetration-testing tool ARTEX, along with CAIRN, CyberStrikeAI, HexStrike AI and PentAGI. In one case, GTIG observed attackers using PentAGI together with proxy services to target the aerospace sector.

The users are not limited to one type of actor. "We have observed several threat actors separately experimenting with use of these frameworks and harnesses, including financially motivated, as well as suspected espionage threat actors with a PRC nexus," Google's researchers wrote. They added that, "given the low barrier to entry, ease of integrating new exploits, and massive scaling potential, we believe it is highly likely that threat actors will continue to use these frameworks in campaigns targeting organisations globally."

The technique at the centre of the case is not new in itself. Patch diffing — reading an update to find the bug it repairs — is long-standing practice among both defenders and attackers. What Google reports here is that the whole chain, from reading the patch to running the exploit against a live target, ran on agents rather than people. That compresses the window between a vendor shipping a fix and an attacker having something that works against systems which have not yet applied it.

Google did not say how long the staging server had been operating, how many organisations were affected, or which software products the framework analysed. The report did not name the operator beyond describing them as Russian-speaking and suspected.

Sources

  1. theregister.comhttps://theregister.com/research/2026/11/16/google-ai-agent-turned-a-security-patch-into-a-working-exploit-with-no-human-help/5309926
  2. thehackernews.comhttps://thehackernews.com/2026/11/google-finds-hackers-using-ai-agents-to.html