Task · CTF
AI for CTF reversing challenges
For CTF reversing, use an LLM to read the check, explain the transform and write a solver. Check the answer by running it in the binary.
Rules
Read the event's rules on AI assistance before you start.
Workflow
- Find the check from the success and failure strings and their cross-references.
- Decompile the check and its callees, and rename what you already understand.
- Ask what happens to each input byte.
- Ask for the inverse as Python you can read.
- Run it. The binary is the judge, not the model.
Prompt
Adapted from the ida-pro-mcp README, for any MCP server.
Analyze this crackme with the MCP tools. Comment and rename as you go. Use int_convert for number bases. Do not brute force. Write report.md and the password.
Tools
| Tool | Why it suits CTFs |
|---|---|
| GhidraMCP | Free with Ghidra. The agent decompiles and renames |
| IDA MCP / ida-pro-mcp | The prompt above comes from its README |
| Binary Ninja MCP | Its README links a CTF demo |
| r2ai and r2copilot | r2ai in the shell, r2copilot for CTFs |
Pitfalls
- A 32-bit little-endian constant read backwards breaks the solver.
- A right shift changes meaning when the value is signed.
- Check loop bounds in the disassembly, and use the tool's number converter.
On a pwn challenge, a model can read the bug and the protections. This site does not cover writing the exploit.
Write-up
After the event, hand the model your notes. With 943,718 output tokens, Kimi K3.1 Pro can draft the write-up in one reply.
Scope
Authorized work only, on your own systems, a scoped engagement, a CTF or a lab. No exploits, payloads, malware or detection-evasion steps.