Skip to content
ComputerBase·Indexed by this site · Originally published· Event →· Its place in the scan record →

Google threat report: AI agent automatically generates exploits from security patches

Autonome Cyberangriffe: KI-Agenten erstellen aus Sicherheitspatches Exploits

AI summary

In its Q4 2026 AI threat report, Google's Threat Intelligence Group (GTIG) and Mandiant describe a framework named exploit_agent that automatically compares code before and after common web application patches, generates working exploits, and executes them automatically; researchers say this is the first observation of such a system requiring no human intervention throughout attack development and execution. The framework runs on a publicly accessible staging server and can connect to external and local language models; because the tooling is in Russian, Google suspects the operator is a Russian speaker, and telemetry data shows targets concentrated on French media, entertainment, and education websites. The report also lists autonomous attack frameworks such as ARTEX, HexStrike AI, and PentAGI, and mentions that a September PaperCut server attack exploited two vulnerabilities 4 days after disclosure. Google warns that the window from patch release to exploit availability is shrinking, and its response includes multi-level safety mechanisms within models, banning abusive accounts and projects, and automated detection and response systems.

Read originalcomputerbase.de