Best LLM for reverse engineering

Short answer: for a reverse engineering agent driving Ghidra MCP or IDA Pro MCP, use a fast model that calls tools and doesn’t refuse. In our test on November 9, 2026, z-ai/glm-5.3-flash-uncensored (Pro) and orca/orcacyber-minimax-m3.1-flash (any plan) both produced a passing serial for a crackme from its Ghidra pseudocode, and both de-obfuscated a PowerShell sample correctly. They start answering in about 2 seconds. Claude Sonnet 5.5 refused the crackme. All of them run on one OpenAI-compatible API, https://obscurify.ai/v1, and in Claude Code.

Reverse engineering is where mainstream assistants refuse most often, even for a binary you wrote yourself or a CTF challenge you’re meant to solve. An agent that stops halfway through a session wastes the work it already did. This page covers which models finish the job, what an RE agent needs from a model, and how to set one up.

Authorized work only. Reverse your own binaries, CTF challenges, malware samples you are analyzing, and software you are licensed or legally entitled to examine. Obscurify.ai’s terms prohibit writing malware and breaking into systems you aren’t authorized to access.

Which model should a reverse engineering agent use?

Two tasks from our cybersecurity LLM test are reverse engineering work. In the crackme task the model gets the serial check as Ghidra pseudocode, explains it, and gives a serial. We count it right only when that serial passes the check function. In the de-obfuscation task the model decodes an obfuscated PowerShell sample we wrote, which is harmless, and we check the decoded script for its marker.

Model idCrackme serialDe-obfuscationFirst tokenPlan
z-ai/glm-5.3-flash-uncensoredPassedCorrect1.7 sPro
orca/orcacyber-minimax-m3.1-flashPassedCorrect2.1 sAny
orca/orcacyber-zero-1.5PassedCorrect1.3 sAny
orca/orcacyber-qwen3.8-flash-nextPassedCorrect1.9 sAny
deepseek/deepseek-v4.1-flash-1017-uncensored:thinkingPassedCorrect11.0 sPro
heabsy/cyberheabsyPartly rightCorrect0.7 sPro
openai/gpt-6.1-solPassedCorrect14.3 sCasual or Pro
google/gemini-3.8-flashPassedCorrect12.2 sCasual or Pro
anthropic/claude-sonnet-5.5RefusedCorrect2.7 sCasual or Pro

First token is the median time to the first word of an answer across the test’s 10 tasks. Each model got one try per task, so treat this as a small test, not a ranking. Current prices and context sizes for the security-tuned models:

What an RE agent needs from a model

For the hardest single problem, such as a custom VM or a heavily obfuscated check, a thinking model is worth the wait.

Set it up

  1. Create an account: a random 20-character code, no email. Add credit and copy your API key from the account page.
  2. Point your coding agent at Obscurify.ai: Claude Code, OpenCode or Cline.
  3. Add the reverse engineering tools: Ghidra MCP, IDA Pro MCP or Reverse Engineer Anything.

We ran the agent half end to end on November 8, 2026. OpenCode on a decensored DeepSeek model reversed a small crackme we compiled ourselves: it read the password from the binary’s data section and ran the binary to confirm it. The tutorials follow each MCP server’s own README.

Malware analysis

The same models work for analysis: de-obfuscating scripts, explaining what a sample does, and writing YARA rules to detect it. Every uncensored and security-tuned model that returned an answer decoded the PowerShell sample correctly (one, Abliterated Model Large v2, returned nothing in time), and the YARA task results are in the full test. Analysis is allowed. Writing or spreading malware is not, under our terms.

Where your samples go

Prompts and replies aren’t stored on Obscurify.ai’s servers. Your identity is removed before a request reaches the model provider, which sees the text but not who sent it. Models with the ZDR tag go only to zero-data-retention routes, based on the providers’ published policies. The OrcaCyber models don’t have the tag, so their provider’s own policy applies. If a client’s binary must never leave your network, run an open-weights model locally instead.

Reverse engineering LLM questions

What is the best LLM for reverse engineering?

In our test, GLM 5.3 Flash Uncensored and OrcaCyber MiniMax M3.1 Flash both solved a crackme from Ghidra pseudocode and de-obfuscated a PowerShell sample. Both answer within about 2 seconds, which matters in an agent that makes many calls. GLM 5.3 Flash Uncensored needs Pro, and OrcaCyber MiniMax M3.1 Flash works on any plan.

Which model works best with Ghidra MCP?

Any model with tool calling that doesn’t refuse reverse engineering. In our test, OrcaCyber MiniMax M3.1 Flash, OrcaCyber Zero 1.5 and GLM 5.3 Flash Uncensored all solved the crackme and start answering in about 2 seconds or less. Run them from Claude Code, OpenCode or Cline with Ghidra MCP added as a tool server.

Does Claude do reverse engineering?

Claude Sonnet 5.5 refused our crackme task, even though the prompt said the binary was a CTF challenge. It did answer the de-obfuscation task. You can still use Claude Code as the agent and run an open model behind it through Obscurify.ai.

Is it legal to use AI for reverse engineering?

That depends on where you are and on the software’s license. Reversing your own code, CTF challenges and malware samples for analysis is common, accepted work. If you’re unsure about a commercial program, check its license and the law where you live. Our terms forbid writing malware and accessing systems without authorization.

Do I need an email to sign up?

No. An account is a random 20-character code, and you can pay by card, Monero or Bitcoin.

Last updated 2026-11-18. Plan prices and model lists on this page are read live from the site.

Try it now: guest chat needs no account, no email and no cookies. When you want more models, an account is just a random 20-character code.

Get an account code See pricing
Get an account code