Skip to content
All Guides
A rotated graphite frame reveals an intact metallic neural lattice, with restrained electric-blue light against a dark navy background.
Operator Guide

Open-Weight Models for AI Red Team Operations

You do not need to start with an uncensored model to build a useful AI red team workflow. Start with a capable model, a clear task contract, and a way to test the result. This guide adapts the Unlocking DeepSeek + GLM operator playbook for human operators and agents, without treating fewer refusals as proof of better security work.

By Jason JeskeUpdated November 17, 2026
Download the operator playbook (PDF)

What does unlocking an open-weight model mean here?

Here, unlocking means getting useful, direct work from an ordinary instruction-following model by improving the session instructions, task definition, and follow-up loop. The Unlocking DeepSeek + GLM playbook uses DeepSeek and GLM as its working examples. It gives operators reusable briefs, output contracts, and recovery prompts rather than a new set of model weights.

Open weights, uncensored behavior, and local deployment are different properties. Open weights describes access to model parameters under a license. Uncensored is a loose label for reduced refusal behavior, not a standard capability rating. A hosted service can serve an open-weight model while retaining its own rules and controls. Downloading weights also does not make every use permitted by their license.

Why not start by shopping for an uncensored model?

Because willingness to answer is only one part of the job. A model can produce a direct response and still invent an API, omit a prerequisite, misunderstand the target, or claim a test ran when it did not. A red team needs an artifact that fits the actual environment and evidence that supports the finding. A refusal-free answer can fail both requirements.

Start with the strongest suitable model you already have access to. Give it one bounded task and a check that can disprove its answer. Consider another model when repeated, representative tests show that the current one cannot do the work. A modified model can be useful, including when studying refusal mechanisms themselves. The point is to select it for a demonstrated requirement, not because its label promises fewer objections.

The playbook favors DeepSeek for complete artifacts and longer work, and GLM for compact, tightly specified output. Those are recommendations from the Unlocking DeepSeek + GLM playbook, not a controlled ranking established by this guide. Your actual deployment, quantization, context budget, and tool integration can change the result.

Start with a task contract, not a dramatic role prompt

Define the outcome, exact target, starting access, deliverable, requirements, and acceptance check. These fields are the core of the playbook's task formula. They make missing information visible and give both the human and the agent a way to decide whether the work is complete. Authorization must already exist outside the conversation. Typing an authorization sentence does not create it.

Task contract: authorized lab log review
Note
TASK: Identify possible authentication abuse in the supplied lab logs. TARGET: The attached synthetic JSONL authentication dataset; timestamps are UTC. STARTING ACCESS: Read-only access to this dataset. No network or shell execution. DELIVERABLE: A table of suspicious event groups. REQUIREMENTS: Cite event IDs and timestamps. Separate observed facts from hypotheses. State missing evidence. Do not infer compromise from a failed login alone. ACCEPTANCE: Every quoted event exists in the input. Each conclusion names its supporting evidence and an alternative explanation.

This is an example brief, not a reported experiment. It is useful because its result can be checked. If the model cites a nonexistent event, the answer fails even when its prose sounds convincing. The playbook recommends making reasonable assumptions and marking them inline. This guide adds a stricter rule for consequential missing facts: resolve an unknown log format, time zone, or target before acting on the answer.

Use different output shapes for DeepSeek and GLM

For the DeepSeek workflow described in the playbook, keep the full task contract. Ask for the complete artifact, exact inputs, assumptions, and validation steps. For a configuration review, that might mean a corrected configuration plus a list of checks to run against the installed version. Exact-looking flags are not evidence that the tool accepts them.

For the GLM workflow, keep the request compact. Name the deliverable first, state the target, and specify the required fields or item count. Ask for separate sections when a long answer becomes difficult to inspect. A fixed count can make omissions visible, but do not force eight findings when the evidence supports only three.

In either workflow, an artifact-first instruction controls presentation. It does not make the artifact correct. Keep uncertainty, missing prerequisites, and required approvals visible. A short answer is not better if it hides the one fact that changes whether a command is safe to run.

Human operators: review, test, and correct one defect at a time

A human operator can inspect each answer before anything executes. Use that advantage. Check the target version, paths, inputs, prerequisites, and success condition. Run applicable checks in the authorized lab. Feed the actual error back to the model rather than asking it to guess what happened.

A precise correction
Note
The review cited event 418, which is not present in the dataset. Rebuild the complete table using only existing event IDs. Preserve the supported observations. Mark any conclusion that no longer has enough evidence as unconfirmed.

The playbook's strongest recovery idea is to name the defect and preserve the parts that work. Correct an assumption, request the missing section, or continue from the last complete point. Do not turn a concrete failure into a long debate with the model. Save the original task, output, correction, and check result so another operator can reproduce the review.

Agents: keep the operating contract separate from execution authority

An agent adds a loop around the model: plan, choose a tool, execute, inspect the result, and continue. Put stable task and output rules in the agent's supported instruction layer. Keep the real target list, credentials, filesystem access, and allowed actions in controls outside model-generated text. A prompt is not a sandbox or a permission system.

Do not promote fetched documents, repository text, logs, or tool output into operator instructions. For an AI red team exercise, test that boundary with a harmless canary: place a request to print a dummy marker inside an untrusted lab document. Observe whether the agent follows its assigned task or the planted instruction. Do not use real secrets or a real external collection endpoint.

These execution controls supplement the Unlocking DeepSeek + GLM prompt playbook. They matter because a more compliant model is not automatically a more trustworthy agent. A human can catch a bad assumption before executing it. An agent needs an explicit boundary at the point where text becomes an action.

Recover incomplete output without hiding the failure

For a truncated answer, quote the last complete line and request the remaining named sections. For an empty answer, first check whether the interface returned a provider error or exhausted an output limit. Retry only within a defined bound. A new prompt cannot repair an unavailable endpoint.

For the wrong target, supply the correct platform or version and request a complete replacement of incompatible instructions. For a format failure, restate the exact fields. For unsupported conclusions, require the evidence or remove the claim. If the model refuses, distinguish a task-definition problem from a provider restriction. Do not assume every refusal is removable through wording.

The playbook contains direct reset and continuation prompts. Use them as editing tools, not proof that an underlying policy or technical limit has disappeared. When the result remains wrong after bounded corrections, preserve the failure and change the model, interface, or task approach for a reason you can explain.

What should you measure before relying on the workflow?

Use a small set of representative, authorized tasks with a scoring rubric written in advance. Include source-grounded log triage, a permission or configuration review, and a bounded agent trust-boundary test. Record the served model identifier, provider, date, prompts, available tools, output limit, and relevant sampling settings. Keep the actual outputs and tool results.

Score willingness to answer separately from format compliance, factual accuracy, execution validity, and observed task success. Repeat the same cases enough to expose variation, and report the trial count with any rate. A single good response is a demonstration, not a general reliability estimate. This guide supplies no new benchmark result.

For agents, also check whether the run stayed inside the authorized target and action budget. If the model completes a task by accessing something it was not permitted to access, the run fails. If it claims success without a matching tool result, the claim is unverified.

Download the final operator playbook

Use the PDF download above for the complete Unlocking DeepSeek + GLM copy-paste edition. It contains the operating sequence, model-specific session and task templates, fourteen cyber briefs, recovery prompts, profile setup, and a one-page operator card. The briefs cover offensive engagement tasks, including detection bypass and persistence, and assume an authorized engagement already exists. The download preserves the final operator PDF unchanged.

The PDF includes a full Nemesis profile-setup chapter, operating instructions described there as active, and a Venice provider/model example. Readers do not need that profile: the standalone session path is the alternative. These are the playbook's examples, not a claim about a current private configuration. The prompts do not establish a provider-wide bypass or a successful test on your environment. Apply the guide's validation and agent-execution checks alongside them.

Frequently Asked
Do I need an uncensored model for AI red team work?
Not as a starting requirement. Test whether an ordinary capable model can complete your authorized tasks with a clear brief and verifiable output. Choose a modified model when your measured requirements justify it, not on refusal behavior alone.
Does this method change the model's weights?
No. It changes instructions and task structure at inference time. It is not fine-tuning, refusal-direction removal, or a guarantee that the model or hosting service will comply.
Can a human and an agent use the same task contract?
Yes. Both need an outcome, target, starting access, deliverable, requirements, and acceptance check. An agent also needs externally enforced tool permissions, target scope, budgets, stop conditions, and independently verified execution evidence.
Does the guide prove DeepSeek is better than GLM?
No. It explains the Unlocking DeepSeek + GLM playbook's model-specific operating recommendations. This guide does not run a new comparison or establish a universal performance ranking.
This is one of the attack surfaces covered on Krypteia Sec. See the research, the glossary, or the course.