T3MP3ST: AI Agents Hunt 0-Day Bugs Autonomously

A new open-source security framework, T3MP3ST, is transforming AI coding agents into autonomous bug hunters. This framework works with existing agents like Claude Code and OpenAI’s Codex, without needing new API keys or cloud infrastructure. T3MP3ST coordinates multiple agent instances, from reconnaissance to exploit and reporting. Users direct the framework at a target, and their existing AI agent becomes the operational brain. The system enforces containment to prevent touching off-scope public hosts. The framework achieved a 90.1% score on the XBOW XBEN suite. It also solved 23 out of 40 tasks on the Cybench benchmark. More notably, it identified 8 out of 10 real vulnerabilities from 2026, which is significant because these bugs were disclosed after the model's training cutoff. Security researchers are calling this a notable development for autonomous red-teaming. This innovation could change how organizations approach cybersecurity and vulnerability detection.