Skip to content
Help Net Security AI· Sinisa Markovic·Indexed by this site · Originally published· Event →· Its place in the scan record →SelectedAttention78

Google finds AI agents writing and launching exploits without human involvement

Google finds AI agent writing and launching exploits with no human in the loop

AI summary

Google Threat Intelligence Group (GTIG) and Mandiant report that attackers have begun having AI agents carry out the full stages of real intrusions; in the past quarter GTIG observed for the first time an LLM framework automatically completing vulnerability discovery and exploitation without human involvement, with the evidence being an exploit_agent multi-agent framework on an exposed server that compared code before and after patches and automatically generated and executed exploits. Open-source agent penetration frameworks such as ARTEX, PentAGI and HexStrike AI have been used by profit-driven threat organizations and suspected China-backed espionage groups, and the FBI has issued a FLASH alert about ARTEX. GTIG counted 22 zero-day exploits in August 2026, higher than the 2025 monthly average of 8; among the LLMJacking incidents handled by Mandiant, some cases had fraudulent costs exceeding 100,000 US dollars in a single day, with the most severe accumulating to 1 million US dollars. GTIG expects that as tools and underlying LLM capabilities improve, human involvement will decrease further and activity will become more covert.

Why it was selected

The report summarizes cases and data of attackers using agent penetration frameworks to carry out real intrusions, and notes that such frameworks currently trade stealth for speed and are more likely to leave detectable traces.

Read originalhelpnetsecurity.com