Money and business in the Middle East.

AI

Google: AI agent turned a security patch into a working exploit

Google says an autonomous multi-agent framework found and exploited flaws with no human in the loop.

· Last verified: 17 Nov 2026 (Google Threat Intelligence Group, Mandiant)

Summary

  • Google says an autonomous multi-agent framework found and exploited flaws with no human in the loop.
  • The exploit_agent toolkit was recovered from an exposed staging server holding stolen victim data.
  • Patch disclosure now hands attackers a map that automated tooling can weaponize within hours.

The latest

A multi-agent framework capable of discovering, weaponizing and launching exploits without human intervention has been recovered from an exposed staging server, Google Threat Intelligence Group and Mandiant said in their Q4 2026 AI threat tracker, published Monday, 16 November 2026. The group described it as the first fully autonomous vulnerability discovery and exploitation it has observed in the wild.

Details

  • How it works: The toolkit, named exploit_agent, was configured to run continuous differential analysis on popular web application software, comparing code before and after security updates to pinpoint likely vulnerabilities, according to GTIG. It then develops and executes the exploit automatically, integrating with several frontier models and locally hosted models.
  • The Russian link: GTIG said the framework relies on Russian-language tooling, including a configuration file carrying comments in Russian. The group did not attribute the operation to a named actor, a state sponsor or a specific criminal group, and published no figure for how many victims or exploits were involved.
  • On the server: Alongside the framework, investigators found data stolen from victim organizations and multiple post-compromise scripts: internal reconnaissance, brute-force password guessing, webshell installation, and routines for removing webshells planted by rival attackers. Victim telemetry suggested the operator was specifically interested in French websites in the media, entertainment and education sectors.
  • The shift: Two months earlier the same group said it had not yet seen fully autonomous attack chains deployed against real-world targets, and framed that as a plausible direction for adversary use of AI. The Q4 tracker treats that threshold as crossed.
  • Frameworks in use: Financially motivated crews and suspected China-linked espionage groups are already using open-source intrusion frameworks including ARTEX, CAIRN, CyberStrikeAI, HexStrike AI and PentAGI, according to GTIG. CAIRN was used against targets including the US aerospace sector, and PentAGI was observed routed through proxy services against aerospace targets.
  • The forecast: GTIG wrote that it had "observed several threat actors separately experimenting with use of these frameworks and harnesses", and assessed it highly likely they will keep using them given the low barrier to entry, ease of integrating new exploits and scaling potential.
  • Patch-to-exploit speed: In the PaperCut print server case, two flaws were chained in late August and researchers along with Huntress built a working AI-assisted exploit within a day of disclosure. Four days after disclosure, an attacker ran a campaign against hundreds of organizations likely using AI, according to GreyNoise.
  • The noise problem: October's patched PHP vulnerabilities produced hundreds of non-functional proof-of-concept exploits, many of them likely AI-generated, the report said. GTIG assessed current agentic tooling as loud and easy to detect, while expecting attackers to grow stealthier and less human-dependent as the tools improve.
  • Stolen credentials: Attackers stole cloud and AI service credentials to run models on victims' accounts, with fraudulent charges exceeding 100,000 dollars in a single day in some cases, GTIG said. One intrusion abused a flaw in the LiteLLM gateway, and a suspected North Korean group altered Claude Code configuration files inside companies so every run loaded a backdoor. Google disabled accounts tied to Gemini misuse.

Between the lines

The differential-analysis design is what changes the defender's math: the security update itself is the input the framework needs. GTIG's own PaperCut and PHP examples show the window between disclosure and weaponization already measured in days. The group's assessment that the tooling is currently noisy is the one advantage defenders still hold, and it says that advantage is expected to erode.

What's next

Watch whether GTIG attributes exploit_agent to a named operator, publishes victim or exploit counts, and whether French media, entertainment and education targets disclose intrusions. The next quarterly AI threat tracker is the reference point.

Source: Google Threat Intelligence Group, Mandiant