Reversing with LLMs · November 2026
AI for reverse engineering
The best AI for reverse engineering is Kimi K3.1 Pro. It reads 1,048,576 tokens, writes up to 943,718, and connects to Ghidra, IDA Pro, Binary Ninja, radare2 and JADX.
Decompiler outputGhidra-style
1undefined4 FUN_00101a40(byte *param_1,uint param_2) 2 3{ 4 ushort uVar1; 5 uint uVar2; 6 uint local_c; 7 8 if (param_2 < 5) { 9 return 0xffffffff;10 }11 if ((*param_1 != 0x5a) || (param_1[1] != 0xa5)) {12 return 0xfffffffe;13 }14 uVar1 = CONCAT11(param_1[2],param_1[3]);15 if (param_2 < (uint)uVar1 + 5) {16 return 0xfffffffd;17 }18 uVar2 = 0;19 for (local_c = 0; local_c < uVar1; local_c = local_c + 1) {20 uVar2 = uVar2 + param_1[local_c + 4] & 0xff;21 }22 if (uVar2 != param_1[(ulong)uVar1 + 4]) {23 return 0xfffffffc;24 }25 return (uint)uVar1;26}
After an LLM passchecked by hand
1/* Checks one serial packet: magic 5A A5, a big-endian 2 16-bit payload length, the payload, then an 8-bit 3 additive checksum. Returns the payload length, or 4 an error code below zero. 5 Likely a UART link layer (guess, medium confidence). */ 6int parse_frame(const uint8_t *frame, uint32_t frame_len) 7{ 8 if (frame_len < FRAME_OVERHEAD) /* 5 bytes */ 9 return ERR_TOO_SHORT; /* -1 */10 if (frame[0] != 0x5a || frame[1] != 0xa5)11 return ERR_BAD_MAGIC; /* -2 */12 13 uint16_t payload_len = (frame[2] << 8) | frame[3];14 if (frame_len < payload_len + FRAME_OVERHEAD)15 return ERR_TRUNCATED; /* -3 */16 17 uint8_t sum = 0;18 for (uint32_t i = 0; i < payload_len; i++)19 sum += frame[4 + i];20 21 if (sum != frame[4 + payload_len])22 return ERR_BAD_CHECKSUM; /* -4 */23 24 return payload_len;25}
FUN_00101a40parse_frameChecks magic, length and checksum, then returns the lengthparam_1frameIndexed as bytes from 0 to the checksumparam_2frame_lenCompared against the header and payload sizeuVar1payload_lenBuilt big-endian from bytes 2 and 3uVar2sumAccumulated modulo 256, compared to the last bytelocal_ciLoop counter
| # | Model | Best for | Context | Max output | Reads | Per 1M in / out |
|---|---|---|---|---|---|---|
| 01 | Kimi K3.1 ProMoonshot AI · 2,780B | Best overall | 1,048,576 | 943,718 | text, images, video | $3.00 / $8.00 |
| 02 | GLM-5.3Z.ai · 753B | Strong for less | 1,048,576 | 131,072 | text | $2.00 / $4.00 |
| 03 | DeepSeek V4.1 Flash 1017DeepSeek · 763B | Best value | 1,048,576 | 131,072 | text, images | $0.24 / $0.90 |
| 04 | GLM-5.3 FlashZ.ai · 320B total, 18B active | Bulk triage | 1,048,576 | 131,072 | text, images | $0.24 / $1.00 |
| 05 | Hy4Tencent · 770B (preview card) | Screenshots for less | 1,048,576 | 65,536 | text, images | $1.60 / $3.20 |
Prices per 1M tokens at Silk Compute, read November 18, 2026. Hy4 is for defensive work only.
GhidraGhidraMCPDecompile and rename in Ghidra.SetupGhidraReVaShort tools for large binaries and firmware.SetupGhidraGhidrAssistExplain, chat and bulk rename.SetupIDA ProIDA MCPHex-Rays server that drives IDA.SetupIDA ProGepettoExplain, comment and rename.SetupIDA ProIDAssistRenames and a knowledge graph.SetupBinary NinjaBinAssist and Binary Ninja MCPA panel, and an MCP server.Setupradare2r2aiDecompile and rename in r2.SetupJADXJADX AI MCPClasses and cross-references in an APK.Setup
| # | Provider | Strongest model for reversing | Context | Prompts stored | Anthropic API | Crypto |
|---|---|---|---|---|---|---|
| 01 | Silk Compute | Kimi K3.1 Pro | 1,048,576 | Prompts and responses are never stored. | Yes | Monero, Bitcoin, Ether, USDT, USDC |
| 02 | abliteration.ai | Abliterated Large v2 (GLM-5.3) | 1,000,000 | Never stored by default | Yes | None listed |
| 03 | Morph | DeepSeek V4.1 Flash Uncensored | 1M | Kept up to 30 days on free and paid tiers | Yes | USDC |
| 04 | Featherless | Qwen 3.8 27B Abliterated | 32K | Not logged | Not listed | None |
| 05 | Unfil.AI | Hermes 3 405B | 128K | Not stored after processing; usage logs with IP kept 30 days | No | Yes |
DecompilerFunction naming and AI decompilingNames, types and constants.ReadMalwareMalware triageWhat a sample does.ReadFirmwareFirmware reversingRegisters and drivers.ReadCTFCTF reversingA solver you can run.ReadConceptsDeobfuscation explainedWhere a model fails.ReadAuthorizedVulnerability researchBugs you may test.Read
Three steps
- Get a key. Silk Compute serves Kimi K3.1 Pro and four other abliterated models from one base URL,
https://api.silkcompute.com. Prepaid credit starts at $5. Create an account. - Point your agent at it. Claude Code, Codex and OpenCode each take a base URL and a model id. The MCP setup page has the config files.
- Add your disassembler's MCP server and ask. Then check every name it gives you against the code.
terminalshell
# Ghidra open with GhidraMCP; Claude Code on Silk Compute
claude mcp add ghidra -- \
uvx --from git+https://github.com/LaurieWired/GhidraMCP \
ghidra-mcp --ghidra-server http://127.0.0.1:8080/
claude "Rename what you can in main and list what you're unsure of."Command from the GhidraMCP README, read November 18, 2026.