PPhying中文
Curated security research, vulnerabilities, advisories and tools for practitioners.

Inside the brain of a hacking robot: Exploring traces | AI Cyber Challenge

theori.io | research | #ai-security | #vulnerability-discovery | #aixcc | #sqlite | #llm-agents | #nginx | #agent-traces
📄 Raw .md⬇️ Download .md🔗 Original ↗

Summary

Walkthroughs of agent trajectories from Theori’s autonomous hacking system: how LLM agents found bugs in SQLite, FreeRDP, Nginx and more during AIxCC practice runs, and what the logs reveal.
Published
Collected

Related coverage

research · theori.io
Building Effective LLM Agents | AI Cyber Challenge
Lessons from building LLM agents for DARPA’s AI Cyber Challenge: decomposing tasks, curating toolsets and adapting to models, applied to vulnerability detection, PoV generation and patching.
research · theori.io
The Frontier Isn’t the Model: Why ‘Good Enough’ Reasoning + Scaffolding Is More Important
Xint researchers argue vulnerability discovery owes less to frontier models than to scaffolding, comparing Mythos’s disclosed results with broadly available models and how to pick a model per task.
research · theori.io
AI Cyber Challenge and Theori's RoboDuck
A look at DARPA’s AI Cyber Challenge and Theori’s open-sourced RoboDuck cyber reasoning system, including its LLM-first design and a pipeline producing POVs without fuzzing or symbolic execution.
research · theori.io
Offensive Security with Large Language Models (1)
Part one of a series on LLMs in offensive security: how projects such as OSS-Fuzz-Gen, Google Project Zero’s Naptime and DARPA’s AIxCC are pushing automated vulnerability detection and exploitation.
research · theori.io
Winning the AIxCC Qualification Round
Theori’s RoboDuck cyber reasoning system took first place in DARPA’s AIxCC qualification round, earning the most achievements and unique bug classes and securing $2M plus a 2025 finals spot.
research · depthfirst.com
Introducing dfs-large1: A New Frontier for Cybersecurity
The release describes a concrete training and evaluation recipe for long-horizon AI code auditing, including repository-scale context, executable validation and explicit pressure against speculative findings.
Back