AI penetration testing
Uncensored AI for penetration testing
AI penetration testing puts a language model inside an authorized test to read tool output, plan checks, review code and draft findings while a tester with written permission decides what runs. PentestLLM supplies that model to security teams and consultancies: uncensored and abliterated builds beside 16 security-tuned ones, by API on one prepaid key.
from openai import OpenAI client = OpenAI(base_url="https://api.pentestllm.ai/v1", api_key="sk-ptl-live-…") reply = client.chat.completions.create( model="auto/security", messages=[{"role": "user", "content": "Explain this advisory and list what to check on an affected host."}],)print(reply.model) # the build that answeredprint(reply.choices[0].message.content)What a security team is buying
- A model API
- PentestLLM is a model API and nothing else. Your testers, your agent and your report template stay where they are.
- The builds
- 16 security-tuned builds, with abliterated and uncensored builds listed beside them. Every build has a page on the models list with its price and context window.
- Two wire formats
- OpenAI-format requests go to
https://api.pentestllm.ai/v1. Anthropic-format requests go to the Messages endpoint on the same key. - Per-token billing
- A prepaid balance from $10 by card. There is no subscription, no seat licence and no renewal.
PentestLLM does not run penetration tests or write reports. Your team or the firm you hire runs the test, and this API is the model their tooling calls.
Why a general model stalls mid-test
A general assistant is trained to decline a wide band of security questions. That band includes what a tester asks on every engagement: what a finding means, how to confirm it on a host in scope, what the proof of concept in a public advisory does. Inside an agent loop one declined step ends the run.
The builds here are abliterated or tuned to answer those questions. What is abliteration explains how the refusal direction is removed from a model’s weights. Refusals are reduced, not removed, which is why the trial further down runs on your own inputs. For what the model does at each stage of a test, read AI pentesting.
What a client will ask about the model
- Prompt text
- Prompts and completions pass through the gateway and are not written anywhere, unless you turn on replay for a key, which keeps that key’s request bodies for 7 days.
- Usage records
- One record per request: the request id, the model, token counts and the charge. Beside it sit SHA-256 fingerprints of the request and the response, never the text.
- Processors
- The host that serves a build receives the request and the response for that call. Check your engagement terms on processors before client data goes into a prompt.
- Build identity
- Once a day every listed model gets the same fixed prompt, and the daily real-model check publishes the result for each one.
- Availability
- The status page shows an hourly API check over the last 30 days, and the same data as public JSON.
The privacy page has the full list of what is stored and who processes it.
Costing an engagement before it starts
Each build has an input price and an output price per million tokens. On 2026-11-20 input prices for security builds start at $0.40 per million tokens, and the table below lists every security build. The engagement cost estimator turns a planned engagement into tokens at live prices and shows how far prepaid credit goes.
A request that would overdraw the balance is refused, not billed, and a request that fails before the model answers is not charged. A non-streamed response carries its settled charge in the x-uplink-charge header, so a script can log cost per finding or per client. The prepaid balance is not a hard cap when auto top-up is on, so put a daily, weekly or monthly spend limit on the key in the console before it goes into an agent.
It runs in the tools your testers use
- Terminal and editor agents
- OpenCode, Claude Code, Cursor, Cline and aider take the base URL and key. The agent guides have the settings for each.
- Kali Linux
- Kali packages
mcp-kali-server, a bridge that lets an agent run commands on a Kali machine. The Kali MCP guide sets it up with every call waiting for approval. - AI system tests
- A red teaming harness needs an attacker model that writes the adversarial prompts. AI red teaming covers that setup.
- One id for routing
- Send
auto/securityand the router picks a fast, balanced or frontier security build per request, at that build’s price with no router fee. - Server-side tools
- Declared in a request, web search ($0.01 per search), page fetch ($0.001 per page fetched), a code sandbox ($0.05 per container-hour, 1-minute minimum) and remote MCP (No tool fee) run on the API side.
Teams weighing a hosted assistant or open weights against an API can read the HackerGPT alternative and WhiteRabbitNeo alternative pages.
How to run a trial
- Load the minimum
- Top up $10 by card and create a key in the console.
- Replay finished work
- Take tool output and code from a closed engagement or a lab, with client identifiers removed, and send it with the model id
auto/security. - Read what came back
- The response names the build that answered. Note whether it answered or declined, and whether its findings match what your tester wrote.
- Pin and compare
- Send the same inputs to two named builds from the models list and compare the answers and the charge.
Written permission comes first
The key is sold for lawful security work on systems you own or have written permission to test, inside the scope of that permission. The acceptable use policy applies to every build, and a key may be suspended until proof of authorization is shown. The rules of engagement builder produces an authorization letter and scope template to fill in before a test starts.
16 security builds, priced per token
Read from the live catalog on 2026-11-20. Prices are US dollars per million tokens, paid from a prepaid balance that starts at $10. The id auto/security has no price of its own: each request is billed at the build that answered.
| Wire formats | |||||
|---|---|---|---|---|---|
| Qwen3.8 27B Cybersecurity | Security | $0.40 | $2.40 | 262K | |
| Uncensored Qwen3.8 27b Cybersecurity | Security | $0.40 | $2.40 | — | |
| Gemma 4 26B A4B Cybersecurity | Security | $0.4224 | $1.3376 | 262K | |
| Gemma 4 26B A4B Cybersecurity (Uncensored) | Security | $0.48 | $1.52 | — | |
| GLM-5.3 Flash Cybersecurity | Security | $0.5093 | $1.6975 | 1049K | |
| GLM-5.3 Flash Cybersecurity (Uncensored) | Security | $0.5093 | $1.6975 | — | |
| Red Team (Flash) | Security | $0.90 | $3.00 | — | |
| Red Team (Pro) | Security | $1.50 | $6.00 | — | |
| OrcaCyber MiniMax M3.1 Flash | Security | $2.00 | $4.00 | 262K | |
| OrcaCyber MiniMax M3.1 Flash (Uncensored) | Security | $2.00 | $4.00 | — | |
| OrcaCyber Qwen3.8 Flash Next (Uncensored) | Security | $2.00 | $4.00 | — | |
| OrcaCyber Qwen3.8 Flash Next | Security | $2.00 | $4.00 | 262K | |
| Uncensored Qwen3.8 Flash Next Cybersecurity | Security | $2.00 | $4.00 | — | |
| Cyber Frontier | Security | $5.25 | $15.75 | 1049K | |
| OrcaCyber Zero 1.5 | Security | $7.50 | $18.75 | 999K | |
| OrcaCyber Zero 1.5 (Uncensored) | Security | $7.50 | $18.75 | — |
Questions
What is AI penetration testing?
A penetration test in which a language model reads tool output, proposes the next check, reviews code and drafts findings. A tester with written permission sets the scope, approves what runs and signs the report.
Is PentestLLM a penetration testing service?
No. PentestLLM is an API that serves language models. Your own team or the firm you hire runs the test, and their agents and scripts call this API for the model.
Is client data stored when it is sent in a prompt?
Prompts and completions pass through the gateway and are not written anywhere, unless you turn on replay for a key, which keeps that key’s request bodies for 7 days. The host that serves the build receives the request and the response for that call.
What does AI penetration testing cost on this API?
It is billed per token from a prepaid balance that starts at $10 by card, with no subscription. On 2026-11-20 input prices for security builds start at $0.40 per million tokens.
Which AI pentest tools work with it?
Any agent or script that takes an OpenAI-compatible base URL and key, or the Anthropic Messages format. The agent guides cover OpenCode, Claude Code, Cursor, Cline, aider, Kali MCP and PentestGPT.
Who runs PentestLLM?
Uplink Switch LLC, a Wyoming limited liability company. It runs the gateway, the console where keys and top-ups live, and the API.
Read next
One key, from $10
PentestLLM is for systems you own or have written permission to test. The acceptable use policy says what is not allowed.