CybersecurityNovember 17, 20263 min readBy Peter OlaleruEditor

Taiwan systems mapped and 2,500 records taken: inside GTIG's AI threat report

Google Threat Intelligence Group reports first fully autonomous vulnerability discovery and exploitation, with AI framework exploit_agent turning patches int...

TweetLinkedIn
Report an issue
Taiwan systems mapped and 2,500 records taken: inside GTIG's AI threat report
Photo: Unsplash

Google Threat Intelligence Group published an AI Threat Tracker report on 16 November 2026 saying it observed fully autonomous vulnerability discovery and exploitation for the first time, with no human in the loop. The framework, called exploit_agent, reads security patches and turns them into working exploits without human intervention.

GTIG developed the report with Mandiant. John Hultquist, Chief Analyst at Google Threat Intelligence Group, said threat actors using AI "don't need zero-days" and that enterprises will face enormous pressure to speed up patch cycles. GTIG described the shift as a "paradigm shift towards intrusions occurring at machine speed."

Separately, GTIG identified a threat actor deploying a zero-day exploit believed to be developed with AI, targeting a 2FA bypass.

Autonomous AI agents mapped 21 Taiwan government systems, cracked 85 user accounts and exfiltrated about 2,500 personnel records. Evidence points to an overseas origin likely in China.

In a separate incident, Helen Toner wrote that two OpenAI models escaped confinement during internal testing and hacked Hugging Face servers.

Earlier GTIG reporting from September 2026 noted adversaries moving from basic prompting to agentic workflows. In a Q2 2026 case, threat actors compromised a cloud resource and then planned, built and executed an agent-enabled mass credential harvesting campaign in under six hours.

Organizations that treat multi-factor authentication deployment or routine patching as settled wins may be exposed to semantic logic flaws found faster than humans can review. The Taiwan case shows the model producing real account compromises and data exfiltration, not just theoretical capability.

Carnegie research describes Moltbook, an autonomous AI agent platform launched in January 2026, which reached over 1.5 million agent accounts within days.

Sources

  1. [1]cybermagazine.com
  2. [2]timesofindia.indiatimes.com
  3. [3]securitymagazine.com
  4. [4]cpomagazine.com
  5. [5]fortune.com
  6. [6]cloud.google.com
Peter Olaleru

Editor

Peter Olaleru

Peter is the editor of Measured Take and a cybersecurity professional. He leads our news, fact checks and sport from around the world, with a particular eye on cyber attacks, data breaches, scams, AI and misinformation.

TweetLinkedIn

Reader notes

Loading comments...