LLM assisted reverse engineering
- ID
- 48452
- Status
- summarized
- Published
- 19 Nov 2026, 6:29 AM
- Fetched
- 20 Nov 2026, 4:26 AM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://www.slowchronicle.com/posts/llm-assisted-reverse-engineering
- Source URL
- https://hnrss.org/best
Summary
- Score
- 7.5
- Created
- 20 Nov 2026, 4:27 AM
- Tags
- Audience
- developersai_agent_usersai_ml_learners
What happened
A self-described two-decade reverse engineer walks through their current workflow: point an agent (DeepSeek Harness with DeepSeek v4.1 Flash, occasionally Codex with Luna or Terra) at a binary, disassemble/decompile via Ghidra through an MCP bridge or faster radare2/kuna CLI tools, and have the agent write every finding to Markdown files as it goes. The write-up stresses that those MD notes serve three jobs — the human's understanding, the agent's memory across context compaction or handoff to another agent, and keeping the agent from circling. When static analysis stalls, the fallback moves are: let the agent debug with breakpoints, inject a logging probe, give it computer use for desktop apps or an API-controllable emulator for firmware, ask for a specifically different approach (e.g. follow strings instead of xrefs), or read the full conversation/thinking trace to spot the blind path.
Why it matters
The transferable part isn't reverse engineering — it's the scaffolding: persist agent findings to Markdown so the work survives compaction and handoff, and read the raw trace when the agent loops. If you run long agent tasks, that's a cheap change you can make this week. Note the text gives no pricing, benchmarks, or accuracy numbers, so treat the tool claims (DeepSeek being 'fast and cheap', Ghidra+MCP being 'best in market') as one practitioner's opinion, not measured results. No Malaysian or SEA angle in this piece.
Discussion angle
Which parts of this workflow are model capability versus harness design? The MD-file notes and the 'try strings instead of xrefs' nudge are harness tricks that work on any long-running agent task — ask the room what scaffolding they've built that outlives a single context window.