One Harness, Three Security Agents: Recursive Language Models in Practice
DAY 2
13:00-
13:40
Most LLM security agents stuff context into a prompt and hope. As the engagement grows the context rots: the model drops an early finding, invents a vulnerability, and burns an unpredictable budget. Everyone rebuilds the same scaffolding, none of it trainable.
Recursive Language Models differ: context lives as variables in a sandboxed REPL, and the model reasons over it in code with recursive sub-calls. rlm-kit is my open-source harness over DSPy’s RLM: declare a task, inherit the sandbox, replayable traces, and reward-free dataset export.
Three agents run on it. cve-reverser turns a disclosed CVE patch into a lab PoC and a Nuclei template. diff-sentry reads a GitHub diff as untrusted data and flags malicious changes before they land. toolscout drives hundreds of MCP tools with a planner holding a few schemas at a time.
The harness, not the model, carries the task structure, so one trained planner transfers across security domains. You work on the task, not on machine learning.
-
Location :
-
Track 3(Room 3)
-
-
Category :
-
Bluebox
-
-
Share :
Speakers
-
Syue Siang Su
シュエシャン・スー
Syue Siang Su is a Security Research Manager at CyCraft Technology and is currently focused on cloud security, web security, and blockchain security. He takes an active role in the cybersecurity community and has delivered speeches at multiple seminars across the globe, including HITCON, HITB, FIRSTCTI, VB, and HackerOne. He still participates in CTF competitions, including SECCON CTF in Japan and HITCON CTF in Taiwan, and has submitted multiple reports to bug bounty programs and open-source projects.