Recoleta
GitHub
Source note

Agentic Vulnerability Reasoning on Windows COM Binaries

Authors
Hwiwon Lee, Jongseong Kim, Lingming Zhang
Source
arXiv
Published
Collection
Software Intelligence
Software SecurityBinary AnalysisAgentic CodingVulnerability DiscoveryWindows Com
Open arXivMarkdown source

Summary

Slyp is an agentic system that finds race-condition bugs in Windows COM binaries and generates debugger-verified C++ proof-of-concept programs. It combines LLM reasoning with binary analysis, COM metadata lookup, and live debugging tools.

Problem

  • Privileged Windows COM services expose RPC-accessible methods to authenticated users, so race conditions can become local privilege escalation bugs.
  • These bugs are hard to find because optimized binaries lose source-level types and symbols, and the analysis must track shared object fields, locks, thread interleavings, and COM vtable dispatch.
  • Existing fuzzers struggle to hit timing windows, and COMRace reports many false positives in this evaluation, with 0.299 F1.

Approach

  • Slyp uses a ReAct-style LLM agent in two stages: discovery and proof-of-concept generation.
  • During discovery, the agent calls MCP binary-exploration tools for decompilation, disassembly, cross-references, call graphs, and COM virtual-call resolution.
  • During PoC generation, the agent calls COM-inspection tools to recover CLSIDs, interfaces, security descriptors, activation patterns, and method signatures.
  • The agent then uses dynamic-debugging tools to compile, run, crash-test, and refine C++ PoCs against a live Windows COM service.
  • The system adds checkpointing, file memory, BM25 retrieval, and stop-time verification so long binary-analysis sessions survive context limits and premature exits.

Results

  • On a benchmark of 20 COM objects covering 40 vulnerability cases, Slyp reaches 0.973 F1 with GPT-5.4.
  • The best reported production-agent baselines reach 0.910 F1 for Codex with GPT-5.4 and 0.724 F1 for Claude Code with Opus 4.6; Slyp leads tested production-agent setups by up to 0.208 F1.
  • Against COMRace++, the paper reports a 3.3x bug-discovery gain, comparing 0.973 F1 for Slyp with 0.299 F1 for COMRace++.
  • For PoC generation, the strongest Slyp configuration verifies 27 of 40 cases, or 67.5%; production coding agents without the COM-inspection and debugging tools verify essentially no cases.
  • The top four Slyp configurations together solve 34 of 40 PoC cases.
  • In production Windows services, Slyp found 28 previously unknown vulnerabilities across 9 COM services; MSRC confirmed all 28, assigned 16 CVEs, and awarded $140,000 in bounties.

Link