No filter AI chat for security work, from your own client

The chat apps that promise no filter still run a model with its own refusals. Connect Open WebUI, a terminal agent or a phone client to an abliterated build with a base URL, a key and a model id, and get a straight answer about the scope you are paid to test.

Published October 25, 2026 · prices as of November 18, 2026

A no filter AI chat app switches off the filter the app itself runs. The one trained into the model stays. For a tester that means the same polite refusal when you paste a finding from an authorized engagement. The fix is to run an abliterated model inside the client you already have: base URL, key, model id, and the refusal is gone at the weights. This is the client-side half of AI with no restrictions.

Three values and a test request

A client that lets you type in a custom OpenAI-compatible endpoint needs the same three things whatever its name. Make the key in the console; it is displayed one time. The id under Model is an abliterated build, meaning the refusal behaviour has been stripped out of its weights, so you do not spend the system prompt fighting it.

Base URL
https://api.pentestllm.ai/v1
Key
sk-ptl-live-…
Made in the console, displayed one time
Model
glm-5.3-flash-abliterated
GLM-5.3 Flash Abliterated, context not in the catalog

Prove the key works from a shell first, before any client is involved. A single request costs a single charge and returns a single answer.

bash
1curl https://api.pentestllm.ai/v1/chat/completions \
2 -H "Authorization: Bearer sk-ptl-live-…" \
3 -H "Content-Type: application/json" \
4 -d '{
5 "model": "glm-5.3-flash-abliterated",
6 "messages": [
7 {"role": "system", "content": "You assist a tester on an authorized web application engagement. The scope is the staging host in the signed statement of work."},
8 {"role": "user", "content": "The search parameter reflects input unencoded. Explain what that lets me demonstrate and how to write it up, two paragraphs."}
9 ],
10 "stream": false
11 }'

The reply uses the ordinary chat completions layout and ends with a usage block, and a response header reports what the request settled for. Any id from the Models page can go in the model field; its per-token price is listed beside it. For the difference between these builds and a jailbreak prompt, see AI with no restrictions.

Each build on the Models page carries a label. Abliterated means the refusal direction was projected out of the weights, and the heretic and derestricted builds count here too. Uncensored is a fine-tune or a merge trained to answer rather than decline. Security builds were tuned on security material and take exploit review, malware triage and tooling questions that a stock model turns away. For an engagement, a Security build is the natural first pick; an Abliterated build is the general-purpose fallback when you want a current base with no reflex to refuse. The pages on uncensored LLMs for pentesting and abliterated models go deeper on both.

Where the refusal lives when the app says there is none

An answer can be blocked at two places. One is outside the model: a classifier reading your message, another reading the reply, or a list of banned words. An app can turn that off and put the fact in its tagline. The other place is the weights themselves. Alignment training taught the model a refusal direction, and that direction fires regardless of what wraps the model. So a chat that advertises no filter still meets a tester asking about a CVE on an in-scope host with a paragraph about responsible behaviour.

Abliteration deals with the second place. It locates the activation direction that distinguishes refused prompts from answered ones and removes that direction from the weights. What comes out knows what it knew before and writes the way it wrote before, minus the trained urge to decline. Uncensored builds arrive at a similar place differently, through a fine-tune on answering completions or a merge with such a model. What is abliteration covers the method in more detail.

A hosted app also chooses the model on your behalf and seldom tells you which one sits behind the interface. Wiring up the client yourself reverses that: you type the id, and the daily real-model check fires one fixed prompt at every id each morning and notes whether the output drifted. Its run on November 18, 2026 marked 140 of the 170 listed models stable.

Nothing clever happens in that check. Each day at the same hour the same prompt goes to every model id. The reply is set against the replies already recorded for that id. Agreement earns the verdict stable, with the date the run has held since; a changed reply earns changed, with its date; silence earns unchecked. Who hosts the weights is never shown. What is shown is whether the id in your client still answers like the model it was yesterday, which a hosted app that hides its model cannot tell you.

Open WebUI

Open WebUI is a common front-end on a tester’s own box or a team server, and it shows the whole catalog in its model picker as soon as the connection is saved.

  1. Open Settings, then Admin, then Connections.
  2. Find Manage OpenAI API Connections and press the plus sign beside it to add one.
  3. In URL enter https://api.pentestllm.ai/v1; in API Key paste the key.
  4. Save, start a new chat, choose glm-5.3-flash-abliterated from the picker and send something.

Put the engagement in the system prompt: the client, the in-scope hosts, the dates of the window, what the rules of engagement forbid. The model’s suggestions then stay inside the same lines you work inside, and a pasted scanner finding gets read against the right target. The Open WebUI guide covers the rest of that front-end.

Terminal agents and phone clients

Coding agents take the same three values, and that is where most testers will want this. OpenCode, Claude Code, Cursor, Cline and aider each have a page here with the exact config: OpenCode, Claude Code, Cursor, Cline, aider. An agent that stops mid-run because its model declined to read an exploit proof of concept is the reason this page exists.

On a phone, an iOS or Android client that accepts a custom OpenAI-compatible endpoint wants nothing more: base URL, key, model id. Requests follow the public chat completions specification, so no plugin is involved. Prefer a script? The docs page carries Python and Node snippets.

The field labels vary by client. Base URL, API host, endpoint and custom provider all want https://api.pentestllm.ai/v1, ending in /v1 with nothing appended. A client asking for the whole path wants https://api.pentestllm.ai/v1/chat/completions instead. Leave any organization or project id box blank. When a client has a button that fetches the model list, click it: a filled list means the URL and the key are both right, and the ids match the Models page.

Other clients with a custom endpoint field take the same values. SillyTavern uses its Custom (OpenAI-compatible) source (the SillyTavern guide on a sister site walks through it). Janitor AI uses a proxy configuration that wants the full chat completions path and a typed model id, with field names from its own proxy quickstart (read October 25, 2026); the Janitor AI proxy guide has those steps.

Scope in the system prompt, and which build for which job

A no-filter model is not a substitute for a frame. Give every client the same system prompt for an engagement: who the client is, the hosts and ranges in scope, the testing window, and the actions the rules of engagement forbid. The model then reads a pasted scanner line or a proof of concept against the right target and keeps its suggestions inside the same boundary you work inside. The rules of engagement tool produces a clean version of that text to paste.

Match the build to the job. A Security build for exploit review, malware triage, CVE reading and detection notes. An Abliterated build on a current base for report prose, scope summaries and code review where you want the latest general knowledge with no refusal reflex. If you would rather not choose, auto/security in the model field selects a security build to fit how demanding the request is, and auto/abliterated does the same across the abliterated builds; the auto-router page explains the routing. The pages on AI for pentesting, bug bounty, CTF and malware analysis go task by task.

The bill for 100 and 1,000 turns

Nothing recurs. You prepay from $10.00 and every request deducts its tokens. Three builds, priced from the catalog at the moment this page was built:

Uncensored models and their prices per million tokens
GLM-5.3 Flash Abliteratedglm-5.3-flash-abliteratedAbliterated$0.4656$1.94
Cydonia 24B v4.3cydonia-24b-v4.3Roleplay$0.484$0.60
Llama 3.3 70B Abliteratedllama-3.3-70b-abliteratedAbliterated$2.80$2.80

Prices per million tokens, read from the live catalog on 2026-11-18.

The example assumes 1,500 input tokens a turn (a system prompt holding the scope, plus the recent history the client sends again every time) and 250 output tokens, roughly three paragraphs of analysis. A long session with a big context setting costs more per turn; terse replies cost less. Every number below is computed at render from the rates in the table.

ModelPer message100 messages1,000 messages
GLM-5.3 Flash Abliterated$0.0012$0.12$1.18
Cydonia 24B v4.3$0.0009$0.09$0.88
Llama 3.3 70B Abliterated$0.0049$0.49$4.90

After a few turns the input side dominates. Because the client resends everything on screen each time, a session near the context limit pushes thousands of tokens in to get hundreds back. You can pull the per-turn cost down two ways: a smaller context setting in the client, or a summary extension that compresses old turns into one paragraph. Neither touches the output, and output is where the three builds in the table differ most in price.

Bonus credit for bigger top-ups and the per-million rate of every model are on the pricing page. A spend cap per key is set in the console, so a client shared across a team stops at the figure you chose.

When something goes wrong

  • A hedge where you wanted an answer. Even a build with no refusal reflex answers better with the frame in front of it. State the engagement and the scope in the system prompt and ask the question again; if a tune still dodges, switch to a plain abliterated build such as glm-5.3-flash-abliterated.
  • Replies that stop short. The client’s maximum response length is lower than the model needed. Raise it there; the API’s max_tokens is whatever the client sent.
  • Two keys in one request. Setting an Authorization header and an x-api-key header to different values earns a 400. Fill in one.
  • The key in the URL. Also a 400. Keys travel only in a header, which keeps them out of server logs and browser history.
  • Prompt retention. Request bodies are not written down unless you turn on the seven-day replay window for a key in the console. Client data you pasted in the morning is not sitting on a disk by the afternoon.
  • An id that is gone. The catalog moves. The Models page lists what is live and the daily check lists which ids answered today; choose from those.
  • A wait before the first token. With stream on, nothing arrives until the model has read the full history, so a long context means a longer wait. Trim the context setting to shorten it.
  • One key, several testers. A key accepts requests from any number of clients, all billed to one balance. Give each person a key of their own from the console, so each has its own cap and can be revoked without touching the others.

If you would rather set up nothing

The hosted apps in the search results have their place. Venice puts chat, image and video generation under one roof and opens its home page on privacy (read October 25, 2026). Other hosted chat apps share Google’s first page for this query (October 24, 2026). For a message in the next half minute with no base URL in sight, a hosted app wins. The prompt box on this site’s home page gives a few demo turns with no account, and the playground runs the same models in a browser tab, so you can try a build before any client is configured.

Wiring up your own client costs one key and three fields. What you get back: the model named by id, the client you already run with its presets intact, a per-token rate printed beside each build, and a daily check confirming that this morning the id you chose still answered as itself.

Questions

The hosted chat apps that rank for that query each set their own terms. A key here needs an account and a prepaid balance from $10.00, and in exchange you choose the exact model and pay per token with nothing recurring. The home page prompt box gives a few demo turns without an account.

An open-weight model with the refusal removed from its weights, served through the same chat completions interface ChatGPT clients already speak. Same client, same prompt format, same streaming, and no jailbreak prompt in front. See AI with no restrictions.

Chat requests are not filtered by topic. The acceptable use policy spells out what is never permitted, and it covers testing systems you are not authorized to test. Image and video prompts involving minors are refused. See the acceptable use policy.

The one running a model you picked on purpose. Choose an id from the Models page by price, context window and label, confirm it on the daily real-model check, then connect it from Open WebUI, a coding agent or a phone client. See every build, priced.

Yes. An Android client with a custom OpenAI-compatible endpoint field takes the base URL, the key and a model id, and sends the same request a desktop client would.

Per token, no monthly fee. The table on this page prices 100 and 1,000 turns on three live builds using catalog rates at render time, and the pricing page covers every model. See per-token prices.

No. A jailbreak is a prompt that argues a filtered model past its training, and it breaks when the model is updated. An abliterated model has the refusal taken out of the weights and needs no trick in the prompt. See abliterated vs uncensored.

Every model above is on the uncensored model list with its live price per million tokens. Per-token pricing sets the whole catalog side by side, and the family directory groups the builds by base model.

Get a keyPrepaid from $10, no subscription.

New models and price changes, once a month.

New uncensored builds with their prices, price moves, and what the daily model check found.

Help