Latest
AI Affairs, home

Tuesday 17 November 2026

News

Google reports AI agent used security patches to build working exploits

Google calls it the first known case of an agent carrying out the process without human involvement; investigators also found stolen data on the operator’s staging server.

Googleplex buildings, outdoor seating and colorful umbrellas beneath a blue sky
Photo: Sylvain Kalache from San Francisco, USA, CC BY 2.0, via Wikimedia Commons (cropped)

Google Threat Intelligence Group said on Monday 16 November that it had identified the first instance it knows of in which an AI agent built working exploits from security patches without human involvement, The Register reported. The case, documented in Google’s fourth-quarter 2026 AI threat tracker, involves a framework that compared software updates, developed exploits and executed them without a person directing each step.

Key points

  • Google describes exploit_agent as an autonomous framework for finding and exploiting vulnerabilities exposed by security updates.
  • Investigators found an exposed staging server containing stolen data and scripts for activity after an intrusion.
  • Server telemetry suggests the operator was interested in French media, entertainment and education websites.

exploit_agent compared code across security updates

Google’s threat hunters found an exposed staging server running exploit_agent, a framework that used several AI agents for vulnerability research and developing attacks. It integrated with frontier and local models and used Russian-language tooling. Google described the operator as suspected to be Russian-speaking.

The framework was configured to keep analysing changes to popular web application software. It compared code from before and after security updates to identify possible vulnerabilities, then developed and ran exploits, according to Google’s account. That takes the work beyond asking a language model for advice on a vulnerability one prompt at a time.

Google calls this the first known instance of that sequence being completed by an AI agent with no human in the loop. Its tracker describes the activity as an advance from the attacker behaviour it documented in the third quarter.

French websites appear in server telemetry

The staging server also held data stolen from victim organisations and scripts for use after gaining access, The Register reported. Those scripts supported internal reconnaissance, brute-force attacks, installation of webshells — files that give an attacker remote access through a website — and removal of webshells left by other attackers.

Telemetry on the server suggests the operator was particularly interested in French websites in media, entertainment and education, according to Google. That description identifies the operator’s apparent targets; the stolen data establishes that victim organisations had already been compromised.

The finding arrives alongside a different response to AI-enabled cyber risks. Tencent and Moonshot AI curbed cyber-hacking capabilities in their open-weight models, AI Affairs previously reported. The UK’s National Cyber Security Centre has also warned that it expects AI-powered ransomware against infrastructure by 2028.

PentAGI appears among other offensive frameworks

Google’s tracker describes other attackers using multi-agent frameworks to conduct intrusions rather than relying on manual prompts. The activity spans financially motivated criminals and suspected espionage groups. Google says several threat actors have separately experimented with the tools, including actors it suspects have a connection to the People’s Republic of China.

The tools named include ARTEX, an open-source penetration-testing tool, as well as CAIRN, CyberStrikeAI, HexStrike AI and PentAGI. In a separate case, Google observed attackers pairing PentAGI with proxy services to target the aerospace sector.

Sources

Topics: Agents, Safety

What should we cover? Tell us

The weekly briefing

The AI Affairs weekly briefing: the week's 10 most important AI stories in business, government and markets, every Monday. We'll email you a link to confirm. Every email has a one-click unsubscribe. Privacy policy