MOUNTAIN VIEWGoogle Threat Intelligence Group (GTIG) and Mandiant have documented cybercriminals using large language model frameworks to discover and exploit software vulnerabilities completely without a human in the loop for the first time. In their November 2026 AI Threat Tracker report, researchers warned that offensive AI has crossed from manual prompting into multi-stage intrusions executed at machine speed.
How does the autonomous exploit_agent framework actually work?
The breakthrough discovery centered on an exposed attacker staging server hosting a multi-agent framework dubbed exploit_agent, wired to both frontier and local AI models alongside stolen victim data, webshells, and brute-forcing scripts. GTIG telemetry showed exploit_agent running automated differential analysis on software patches to weaponize flaws against French media, entertainment, and education websites without human oversight.

Beyond exploit_agent, attackers are actively abusing open-source offensive suites including ARTEX, CAIRN, CyberStrikeAI, HexStrike AI, and PentAGI in live operations. Between late August and early October 2026, GTIG tracked an operator running a customized PentAGI orchestration platform to systematically scan organizations across Asia, while CAIRN struck the U.S. aerospace sector and ARTEX drew an FBI warning after hitting defense, healthcare, and financial firms.
How fast are AI-driven vulnerability exploits accelerating?
Vulnerability disclosures and exploits nearly doubled between January and August 2026, with zero-day exploitation climbing from a 2025 average of eight per month to 11 in 2026 and peaking at 22 in August. Flooding defenders with noise, hundreds of largely AI-generated proof-of-concept exploits also surfaced immediately after two PHP2Shell flaws were disclosed, and GTIG projects sophisticated fully autonomous agentic attacks could emerge within six months.
Adversaries are simultaneously weaponizing developer AI assistants and cloud credentials for profit and espionage. Google observed activity cluster UNC6836 altering Claude Code configurations on compromised systems to deploy the SOMBERMEME backdoor at cryptocurrency firms in North America and Europe, while China-nexus espionage actors tested custom automated frameworks and profit-driven criminals resold stolen corporate model access through underground LLM Mixers.
Why does GTIG's Asia warning matter for Indonesian defenders?
Because PentAGI specifically scanned targets across Asia and automated patch-diffing compresses remediation windows, Indonesian enterprises, banks, and public agencies face sharper pressure to patch internet-facing software rapidly and audit AI coding tools. While autonomous agents still trade stealth for speed and leave detectable forensic traffic for Google SecOps rules, the warning builds on earlier AIGOS News coverage of AI exposing the Java4Shell flaw and Google removing 19,389 influence channels in Indonesia.
