Security research

Uncensored LLM for Security Research: A Field Guide

An uncensored LLM for security research is a model with its refusal behavior removed, so it reads the real sample and answers. For this work, use Kimi K3.1 Pro. It's the best model for security research.

Here's what to hand it, prompts you can paste, and the habits that keep a confident model honest.

A hex dump printed in dark ink on pale paper, with a run of bytes marked in orange

Why do general-purpose models get in the way?

It's 2 a.m. You've pulled a sample off a compromised host, you paste the decompiled function, and the model tells you it can't help with malware. Of course it looks like malware. It is malware. That's the job.

Refusal training is broad and blunt, so it catches the people defending against this stuff too. The problem is common enough that researchers built OR-Bench, an ICML 2025 benchmark with 80,000 prompts, just to measure how often models turn down requests they shouldn't. An abliterated model has that refusal behavior removed, so you spend the night on the sample and not on rephrasing the question.

Which model should you use?

The jobModelEffort
Malware analysis, reverse engineering, vulnerability researchKimi K3.1 Promax
The same work at a lower billGLM-5.3max
Reviewing a large codebase in an agent, all dayDeepSeek V4.1 Flash 1017high
Screen recordingsKimi K3.1 Promax
Screenshots, charts and dashboardsHy4high
Bulk triage of strings, alerts and log linesGLM-5.3 Flashlow

Kimi K3.1 Pro is the best model for this work. GLM-5.3 at max effort is the strong choice for less: Z.ai reports 84.5 on CyberGym, a benchmark of 1,507 real-world vulnerabilities across 188 software projects, run in Claude Code at max effort. That's the top score on the CyberGym row of its own table, for the original model before abliteration.

Both read up to 1M tokens, so long samples and whole codebases fit. Both cost more than DeepSeek and GLM-5.3 Flash, so don't use them for routine volume. Our coding agent guide shows what each model costs. GLM-5.5 is coming soon.

What should you hand it?

No model reads a binary, so decompile first. Paste Ghidra or IDA output, strings and logs.

  1. SampleFrom an incident or your lab
  2. DecompileIn your own sandbox
  3. Kimi K3.1 ProThe best model for this
  4. CheckBy hand, inside your sandbox
  5. ShipA rule or a report

Malware triage

Start here. Ask what the code does, not whether it's bad, and ask for something you can test.

Prompt
Here is a decompiled function from a sample we pulled off a compromised host. Explain what it does, list any network indicators, and write a YARA rule that matches it.

Reverse engineering

Renaming functions and recovering structs is tedious, and that's exactly where this helps. A survey of 48 studies found real potential and plenty of open challenges, so treat every name it suggests as a guess.

Prompt
Here is a stripped function from Ghidra. Suggest names for the function and its variables, recover any structs you can, and mark each guess with how sure you are.

Detection engineering

Have it write the rule, then have it pick the rule apart. The second step is where the false positives show up.

Prompt
Here is the behavior we saw, as process and network events. Write a Sigma rule for it, then list the legitimate software most likely to trigger false positives.

Reviewing your own code

Point an agent at a repo you own. Agents can find real bugs: the CyberGym authors report 34 zero-day vulnerabilities turning up during their work.

Prompt
Review this service for authentication and injection flaws. For each one, give the file, the line, how it could be triggered in a test environment, and the fix.

Incident timelines

With 1M tokens of context, you can paste a lot of logs at once, and a model is happy to read all of them.

Prompt
These are auth, proxy and EDR logs from one host during an incident we're handling. Build a timeline, flag the first sign of compromise, and list what you can't tell from these logs.

How do you set it up?

It's a standard OpenAI or Anthropic request. Change the base URL, use a Silk Compute key, and set the effort.

Python
import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.silkcompute.com",
    api_key=os.environ["SILK_API_KEY"],
)

response = client.chat.completions.create(
    model="kimi-k3.1-pro-uncensored",
    reasoning_effort="max",
    messages=[{"role": "user", "content": """Here is a decompiled function from a sample we pulled off a compromised host. Explain what it does, list any network indicators, and write a YARA rule that matches it."""}],
)

print(response.choices[0].message.content)

To work in a coding agent instead, set it up for Claude Code, Codex or OpenCode. Every model works with tool calling, so the agent can read files and run commands as usual. The setups start on DeepSeek, so switch to Kimi K3.1 Pro at max effort: /model and /effort in Claude Code, /model in Codex, and /models and ctrl+t in OpenCode.

How do you keep the results honest?

A confident answer isn't a correct one. A few habits help:

  • Treat names and claims as hypotheses. Models invent function names and misread control flow. Check anything that matters in a debugger, inside your sandbox, before it goes in a report.
  • Keep samples off your own machine. If an agent can run tools, give it a VM with no credentials, no shared folders, and outbound traffic only to the API. Never detonate anything on the host.
  • Expect strings to talk back. A sample's strings and a repo's comments are untrusted input, and they can carry instructions aimed at the model. An agent with tools might follow them, so keep it read-only where you can.
  • Give it your baseline. It can't tell what's normal on your hosts unless you say.
  • Expect fewer refusals, not zero. An abliterated model can still decline, and it's a different model, not the same one minus the no.

What are the rules?

This guide is for work you're authorized to do: your own systems, your employer's, incidents you're handling, research in your lab. You're responsible for how you use the models, and our Terms say so.

Prompts and model outputs pass through our servers and are not stored. We keep only billing records, such as token counts and cost.