Uncensored LLM for Red Teaming and Pentesting: A Field Guide
5 min read
An uncensored LLM for red teaming is a model with its refusal behavior removed, so it plans attack paths on an authorized engagement instead of asking you to justify the work. Use Kimi K3.1 Pro. It's the best model for red teaming.
You still run the engagement. The model reads the scan output, chains the findings and drafts the report, and you decide what actually gets touched.

Why do general-purpose models get in the way?
You're three hours into an authorized test. You paste the scan output, ask what to try next, and the model wants to talk about ethics. So you rephrase, and rephrase again, with the clock running.
Recon, exploitation and privilege escalation are the job, and LLMs are surprisingly well suited to it, because so much of pentesting is pattern matching. Even a 2026 study of 400 autonomous runs built a one-shot authorization re-prompt into its harness. No refusal survived it, but someone had to build it. With an abliterated model, you stop arguing and get back to testing.
Which model should you use?
Different phases want different models. Here's how we'd split it:
| Phase | What the model does | Model | Effort |
|---|---|---|---|
| Recon | Reads scan output and ranks what looks interesting | GLM-5.3 Flash | low |
| Attack paths | Chains findings into likely paths, ranked by impact | Kimi K3.1 Pro | max |
| Verification | Writes read-only checks for in-scope hosts | Kimi K3.1 Pro | max |
| Reporting | Turns your notes into findings and fixes | DeepSeek V4.1 Flash 1017 | high |
Kimi K3.1 Pro is the best model for red teaming. GLM-5.3 at max effort is the strong choice for less: Z.ai reports 84.5 on CyberGym, the top score on that row of its own table, and says it more than doubles GLM-5.2 on exploitation benchmarks. Those are its numbers for the original model, before abliteration.Kimi K3.1 Pro is there for the hardest workloads, and GLM-5.5 is coming soon.
The model changes the result. In that 400-run study, full exploitation ranged from 25 to 85 of 100 runs across four models. The lowest, a small 14B open model, most often failed by stopping too early.
What does each phase look like?
- ScopeWritten, and enforced outside the model
- ReconScan output in, ranked list out
- Attack pathsKimi K3.1 Pro
- VerifyRead-only checks, you approve
- ReportFindings and fixes
Recon
Tools produce far more output than anyone wants to read. Let the model read it all and tell you where to look first.
We have written authorization to test staging.example.com. Given this nmap output, list the most likely attack paths and the checks to run next, ranked by impact.Attack paths
This is where the best model pays off. Three medium findings often add up to one serious path, and that's what a client actually needs to hear about.
Here are our findings so far on the in-scope hosts. Chain them into likely attack paths, rank them by impact, and say what evidence would confirm or rule out each one.Verification
Ask for checks that prove the point without changing anything. Read the script before you run it. The firewall is still what counts.
Write a small script that checks our in-scope hosts for the misconfiguration in finding 3. It must be read-only, and it must refuse any host that isn't in scope.txt.Reporting
Nobody got into this line of work to write reports. With 1M tokens of context, it can read all your notes at once.
Turn these notes into findings for the client report: title, severity with your reasoning, evidence, steps for their team to reproduce it, and a fix.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.
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 are our findings so far on the in-scope hosts. Chain them into likely attack paths, rank them by impact, and say what evidence would confirm or rule out each one."""}],
)
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 an agent in scope?
A model will never care about your rules of engagement the way you do. So build them in:
- Enforce scope outside the model. Put it in the system prompt, then back it with an allowlist at the network layer. The firewall is what actually counts.
- Approve anything that can change or break a target. That includes logins and heavy scans. A person should say yes to each one.
- Log every command. You'll want them for the report, and the client may ask.
- Treat target output as untrusted. A banner or an error page can carry instructions aimed at the model, so don't let it act on what it reads without you.
- Run important checks twice. The 400-run study held prompt, target and setup constant, and the outcomes still varied from run to run.
- Check that a refusal is a refusal. An earlier draft of that study logged API overload errors as safety refusals. Read the status code before you change models.
What are the rules?
This guide is for authorized work: a signed engagement with a written scope, your own systems, or a lab. You're responsible for staying in scope, and our Terms make you responsible for following the laws that apply to you.
Prompts and model outputs pass through our servers and are not stored. We keep only billing records, such as token counts and cost.