Red teaming
Uncensored AI for cybersecurity.
4 security builds that answer as an analyst, 5 general uncensored builds you steer yourself. One OpenAI-compatible key, priced per token, in Strix, HackingBuddyGPT and CAI.
Picking a model
- Red Team (Pro)
- $13.80 per 1,000 triage turnstriage, CVE write-ups and reports with no system prompt of your own
- GLM-5.3 Uncensored
- $2.37 per 1,000 triage turnsthe budget pick for agent loops that send thousands of turns
- CyberHeabsy
- 1024K contextfor a tool that reads long logs or a whole repository
A triage turn
- Input
- 6,000 tokensa log excerpt or a finding
- Output
- 800 tokensa short analysis
- Prices as of
- 2026-10-27
Security builds
| CyberHeabsy | Security | $5.25 | $15.75 |
| GLM-5.3 Flash Cybersecurity | Security | $0.525 | $1.75 |
| Red Team (Flash) | Security | $0.90 | $3.00 |
| Red Team (Pro) | Security | $1.50 | $6.00 |
General uncensored builds
| GLM-5.3 Uncensored | Uncensored | $0.282 | $0.846 |
| DeepSeek V4.1 Flash Abliterated | Abliterated | $0.8488 | $1.6975 |
| GLM-5.3 Abliterated | Abliterated | $2.25 | $4.50 |
| DeepSeek V4 Pro Uncensored | Uncensored | $1.00 | $2.00 |
| Kimi K3 Uncensored | Uncensored | $5.745 | $28.68 |
Why these builds
The complaint in every security thread is the same: the stock frontier models decline the question, and the vendors gate security work behind verification programs. The builds on this page answer it. The security ids answer as an analyst would, with the identifiers, the evidence and the fix; the uncensored ids answer in whatever shape your prompt asks for. Each model page shows the daily check, and the key works in the agents below today. The service is for the assessment of systems the customer is authorized to test; the acceptable use policy has the line.
One request
curl https://api.uncensoredllm.ai/v1/chat/completions \ -H "Authorization: Bearer sk-ullm-live-…" \ -H "Content-Type: application/json" \ -d '{ "model": "uplink-red-team-pro", "messages": [{"role": "user", "content": "Write a Sigma rule that detects CVE-2024-3094 exploitation attempts and explain the detection logic."}] }'Agent tools
Each of these open-source agents takes an OpenAI-compatible endpoint. The settings are the ones their READMEs name; swap the model id for any row above.
# Strix (usestrix/strix): any OpenAI-compatible endpointexport STRIX_LLM="openai/uplink-red-team-pro"export LLM_API_KEY="sk-ullm-live-…"export LLM_API_BASE="https://api.uncensoredllm.ai/v1"strix --target ./app# HackingBuddyGPT (ipa-lab/hackingBuddyGPT)python -m hackingBuddyGPT.cli LinuxPrivesc \ --llm.api_key "sk-ullm-live-…" \ --llm.api_url "https://api.uncensoredllm.ai/v1" \ --llm.model "uplink-red-team-pro"# CAI (aliasrobotics/cai): .envOPENAI_API_KEY="sk-ullm-live-…"OPENAI_BASE_URL="https://api.uncensoredllm.ai/v1"CAI_MODEL="uplink-red-team-pro"Daily check
| Model | Verdict | Context |
|---|---|---|
| CyberHeabsy | Steady since Oct 24 | 1024K |
| GLM-5.3 Flash Cybersecurity | Steady since Oct 23 | 1024K |
| Red Team (Flash) | Not checked yet | not published |
| Red Team (Pro) | Not checked yet | not published |
| GLM-5.3 Uncensored | Steady since Oct 12 | not published |
| DeepSeek V4.1 Flash Abliterated | Steady since Oct 19 | not published |
| GLM-5.3 Abliterated | Steady since Oct 12 | not published |
| DeepSeek V4 Pro Uncensored | Steady since Oct 12 | not published |
| Kimi K3 Uncensored | Steady since Oct 22 | 256K |
Refusals
A reply that is a refusal is not billed: the gateway detects the pattern and refunds the request, and the refund shows in the usage log with its reason. The uncensored builds rarely trigger it; if one does on a prompt that is within policy, send the request id to support.
Questions
One that answers the security question and fits the budget. Red Team (Pro) (uplink-red-team-pro) leads here at $1.50 in and $6.00 out per million tokens, with a security-analyst instruction built into the route. GLM-5.3 Uncensored (glm-5.3-uncensored) is the budget pick at $0.282 in and $0.846 out. The agent tools above take any of them through the same base URL.
For a tool that reads long logs or a whole repository, the widest context window: CyberHeabsy at 1024K tokens. For a chat analyst that cites CVEs and writes the finding, the security builds. For a model you steer with your own system prompt, the general uncensored builds in the second table.
The 5 general uncensored builds here answer what a stock build of the same weights declines; each model page shows the daily real-model check. The security builds carry a security-analyst instruction ahead of your prompt, so the answer arrives as a finding with its evidence, its proof of concept and its fix.
A local 7B or 8B security model fits a laptop GPU and knows less than a 2026 frontier model. The builds here are full-size open-weight models served by API, so there is no GPU to buy and the id goes into the same tools a local model would. GLM-5.3 Uncensored is $0.282 per million input tokens.
Yes, and the open-source agents expect an OpenAI-compatible endpoint: Strix takes LLM_API_BASE, HackingBuddyGPT takes --llm.api_url, CAI takes OPENAI_BASE_URL. Set each to https://api.uncensoredllm.ai/v1 with your key and a model id from the table.
Two different things share the phrase. Red-teaming an LLM means attacking a model or an AI app to find its failures (prompt injection, data leaks); tools like Promptfoo and DeepTeam do that. This page is the other direction: LLMs for red teamers, models that help a security engineer assess systems the customer is authorized to test.
The model answers; the operator decides what to run. Every request here is billed to a prepaid key, the daily check shows which model answered, and the service is for assessment of systems the customer is authorized to test, as the acceptable use policy says.