---
title: "Private AI: four routes and a no-retention API"
canonical: https://pentestllm.ai/blog/private-ai
generated: 2026-11-18T03:45:57.834Z
source: https://pentestllm.ai/blog/private-ai
base_url: https://api.pentestllm.ai/v1
---

# Private AI for security teams: four routes, and an API that writes down no prompts

Private AI puts you in charge of who reads your prompts and what survives the request. For a security team that is scope letters, client source and unreported findings. Four routes compared: an API that keeps no prompt text, a model on your own hardware, an enterprise private cloud, and a chat app that promises no logs, with open-weight builds you can bring in-house later.

Published October 27, 2026 · prices as of November 17, 2026

Private AI means you choose who reads your prompts and what survives the request. A security team gets there four ways: a model on its own machine, a model in the organisation’s private cloud, a chat app that promises to keep nothing, or an API that writes down no prompt or reply text. This page sets the four side by side, then follows the tooling path through the API route, on open-weight builds that can move to your own hardware when an engagement requires it.

## A working definition of private AI

Private AI is any setup where the user, not the vendor, decides who reads the prompts, what is retained and for how long. The vendor pages on Google’s first page for this query on October 27, 2026 (VMware, Cloudera, Teradata, HPE) frame it as models, data and compute held within one organisation. The top consumer result, a chat app, frames it as conversations that nobody else stores. Neither is wrong. The disagreement is about where the model runs.

Any setup can be measured with three questions:

- **Who reads the text?** The machine that runs the model must see the prompt to answer it. It is your laptop, your organisation’s cluster, or a third party’s server.
- **What is retained?** The text of prompts and replies, usage rows, account details, logs, and how long each of them lives.
- **Who holds the weights?** A closed model exists only on its owner’s servers. Open weights can be carried to your own hardware the day a client’s terms demand it.

## Why a security team cares

A tester’s prompts carry the material every engagement contract is careful about: scope documents, client source code, hostnames and credentials from a target in scope, findings not yet reported. A stored prompt can surface in a breach, be read by an employee, or be produced when a court orders it. Nothing can be produced from a prompt that was never kept.

Anyone building tooling has a second concern. An internal tool that forwards client text to a model inherits the retention policy of the service behind it. If your engagement letter promises the client that nothing is kept, the API behind your tool must keep nothing, and that has to be in writing.

Regulated clients add a third. HIPAA, GDPR and their relatives ask where the data travelled and who processed it. This page claims no compliance status for any option. It sets out the facts each option gives you when that question arrives.

## The four options

| Route | Who reads the prompt | What survives | How big a model | What you pay for |
| --- | --- | --- | --- | --- |
| On your own machine (Ollama, llama.cpp, LM Studio) | Your machine alone | Whatever your front-end logs | Whatever your GPU memory holds | Hardware and electricity |
| Enterprise private cloud | Your organisation’s cluster | Your organisation’s policy | Large, on GPUs you own or lease | Vendor contract plus infrastructure |
| Consumer chat app promising no logs | The app and whoever hosts its model | Whatever the app’s policy says | The app decides | A free tier, then a subscription |
| An API with no retention (this one) | The host running the model, while the call is in flight | One usage row plus fingerprints, no bodies | Every catalog build, 70B and up | Per token, prepaid from $10 |

**Local** is the most private option and the most constrained. A recent laptop runs an 8B; the big open models want far more GPU memory than any consumer card carries. The [guide to uncensored models on Ollama](https://pentestllm.ai/guides/ollama-uncensored-models) walks the local route.

**The enterprise private cloud** is what the vendor pages mean. It suits an organisation that already has security staff, a procurement desk and money for both.

**Chat apps** start fastest and verify worst: you read the policy and take it on trust. Three of them are compared, with dates, on the [chatbot page](https://pentestllm.ai/blog/private-ai-chatbot).

**The API route** sits between the first two. Large models with no hardware to buy, from a provider that states in writing what it retains. The remainder of this page takes that route apart line by line, with the builds a security team would pick.

## What this API writes down, and what it never does

Taken from the [privacy page](https://pentestllm.ai/privacy), last updated 2026-10-27, read for this post on October 27, 2026:

- Prompt and reply text: Not stored Retained for 7 days only when you switch replay on for a key. Otherwise the text crosses the gateway and nothing is written.
- Each request: One usage record The request id, the model, token counts by class with their source, and what was charged.
- Fingerprints: SHA-256, never text A SHA-256 of the request and one of the response, which let a signed receipt prove the charge; plus one of each large system prompt, kept 7 days, for caching prompts shared across accounts.
- Account: Email, balance, ledger Your sign-in address, what checkout reported you paid, and every change to the balance. The card number stays with the processor.
- Keys: A salted hash Displayed once, never kept. What is kept is the hash with the key’s name, status and limits.
- First visit: Referrer and landing page Where you came from, the campaign tags on the link and the page you landed on, stored with the account at your first sign-in.
- Who reads the text: The model host Hosting providers contracted by the operator run the models, and each one sees the request and the response for the calls it serves.

The final row is the ceiling on every hosted option. The text must reach the machine running the weights, or there is no answer. The gateway’s contribution is that no copy is made in passing, and that the one record kept is numbers and hashes open to audit, never prose.

The privacy page says nothing about model training, so neither does this post. Chat requests are not filtered by topic; the [acceptable use policy](https://pentestllm.ai/acceptable-use) names the things that are never allowed, testing systems you have no authorization for among them.

## Weights you can walk away with

The case for an open-weight build in private work is the exit door. Every build in the table has its own weights on Hugging Face, so the id you call today is a repository you can pull onto your own server the day a client’s terms require it. Card facts come from the model cards; prices come from the live catalog when this page renders.

| Build | Repository | Licence per card | 10,000 calls by API |
| --- | --- | --- | --- |
| [NeuralDaredevil 8B Abliterated](https://pentestllm.ai/models/neuraldaredevil-8b-abliterated) Daredevil 8B Abliterated (Llama 3 8B) | on Hugging Face | llama3 | $40.48 |
| [Gemma 4 26B-A4B ChimeraX Heretic](https://pentestllm.ai/models/gemma-4-26b-a4b-chimerax-heretic) Chimera-X 26B-A4B (Gemma 4 26B-A4B) | on Hugging Face | apache-2.0 | $14.24 |
| [Qwen3.8 27B Obliterated](https://pentestllm.ai/models/qwen3.8-27b-obliterated) Qwen3.8 27B | on Hugging Face | apache-2.0 | $33.45 |
| [Llama 3.3 70B Abliterated](https://pentestllm.ai/models/llama-3.3-70b-abliterated) Llama 3.3 70B Instruct | on Hugging Face | llama3.3 | $64.40 |

The job: 10,000 scanner findings triaged, 2,000 tokens in and 300 out per call, at the rates below.

| Model | Build | Input $/M | Output $/M | Context | Wire formats |
| --- | --- | --- | --- | --- | --- |
| [NeuralDaredevil 8B Abliterated](https://pentestllm.ai/models/neuraldaredevil-8b-abliterated) neuraldaredevil-8b-abliterated | Abliterated | $1.76 | $1.76 | 8K | openai |
| [Gemma 4 26B-A4B ChimeraX Heretic](https://pentestllm.ai/models/gemma-4-26b-a4b-chimerax-heretic) gemma-4-26b-a4b-chimerax-heretic | Abliterated | $0.484 | $1.52 | 262K | openai |
| [Qwen3.8 27B Obliterated](https://pentestllm.ai/models/qwen3.8-27b-obliterated) qwen3.8-27b-obliterated | Abliterated | $0.582 | $7.27 | 524K | openai |
| [Llama 3.3 70B Abliterated](https://pentestllm.ai/models/llama-3.3-70b-abliterated) llama-3.3-70b-abliterated | Abliterated | $2.80 | $2.80 | 33K | openai |

Prices per million tokens, read from the live catalog on 2026-11-17.

Not every build offers that exit. DeepSeek V4.1 Flash Abliterated is served here with the refusal removed, but the public weights are the stock base, deepseek-ai/DeepSeek-V4.1-Flash (MIT, 552B backbone, 8B active per token at prefill and 16B at decode). Bringing that job in-house gets you the base model, not the edit.

On November 17, 2026 the daily check at /proof sent its fixed prompt to every listed model, and 143 of 170 answered stable. That check is how you know the id you called this week matches the weights you would download: a build whose answer shifts under its own name gets flagged. The [proof page](https://pentestllm.ai/proof) shows the verdict per model.

## Setting it up as the person who writes the tooling

1. **Classify the data first.** Material that may never leave your hardware under the engagement terms goes to a local model. The rest can go to an API whose retention you have read.
2. **Pick a build with published weights**, from the table or the [models page](https://pentestllm.ai/models), to keep the exit open.
3. **Make a key** in the [console](https://app.pentestllm.ai/console) with a monthly spend cap. Do not turn on replay; it stays off until you choose otherwise.
4. **Point the SDK at this base URL.** That is the only change the OpenAI SDK needs. Ollama answers the same chat-completions shape on localhost (its [notes on OpenAI compatibility](https://docs.ollama.com/api/openai-compatibility) describe it), so relocating the job is one edited line.
5. **Mind your own logs.** An API that keeps no bodies buys you nothing if your tool writes each prompt to its own log file.

```
from openai import OpenAI

# Today: the hosted build. Later: the same weights on your own server.
client = OpenAI(base_url="https://api.pentestllm.ai/v1", api_key="sk-ptl-live-…")
# client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama")

label = client.chat.completions.create(
    model="neuraldaredevil-8b-abliterated",
    messages=[
        {"role": "system", "content": "Triage this scanner finding from an authorized test as critical, high, medium, low or informational. One word."},
        {"role": "user", "content": open("finding-0001.txt").read()},
    ],
    max_tokens=5,
)
print(label.choices[0].message.content)
```

The commented-out line is the migration. Download the same weights to a server you control, point the client at it, and nothing else in the code moves; the model string becomes the name your server gave the download.

## Mistakes that undo the privacy

- **Hearing "private" as "unseen".** Whatever runs a hosted model reads the prompt. The question worth asking is what remains afterwards.
- **Enabling replay for a debugging session and forgetting it.** Replay holds that key’s request bodies for 7 days. Debug on a throwaway key and revoke it when you are done.
- **Pasting credentials.** Passwords, API keys and third parties’ personal data belong in no prompt on any service, even when they came off an in-scope host. Redact before the call.
- **Accepting a slogan in place of a list.** "Your privacy matters to us" is not a retention statement. A row-by-row account of what is stored is.
- **Standardising on a closed model.** Unpublished weights mean the local option is closed forever, however sensitive the engagement turns out to be.
- **One key for every tool.** A key per tool keeps caps and revocation separate, and the usage rows then tell you which tool sent what.

## Questions

### Which AI can I use privately?

A local open-weight model through Ollama, llama.cpp or LM Studio for anything that must stay on your machine. For bigger models, a service whose retention you have read: on this key, prompt and reply bodies are not stored unless replay is on, and a key costs an email and a top-up from $10. See what the privacy page lists.

### Does a private ChatGPT exist?

ChatGPT lives on OpenAI’s servers under OpenAI’s data terms and nowhere else. The open-weight path is GPT-OSS 120B, OpenAI’s own release under Apache-2.0; an uncensored build of it is listed on this key and reachable with the same chat-completions call. See the models page.

### What does private AI cost?

Local: the hardware and the electricity. Enterprise private cloud: a vendor quote plus GPUs. Consumer apps: their own published plans. This API: per token, nothing recurring; the table on this page costs a 10,000-finding triage job on four builds at live catalog rates. See per-token prices.

### Can I build my own private AI?

Yes. Pull an open-weight model, run it under Ollama or llama.cpp, and put Open WebUI or similar in front. When a model is too big for your card, call the same build by API with replay left off, and move it home later if its weights are public. See Open WebUI on this key.

### Is private AI free?

Once the hardware is paid for, a local model costs nothing more. On this site the playground gives a few free turns a day without an account on the builds marked free; beyond those you prepay a key and pay per token. See try the playground.

### Is Private AI the name of a company?

A PII-redaction company was called Private AI; a search ad seen on 27 October 2026 said it now trades as Limina. Here the phrase is generic: AI where you decide who sees the data.

### Does a private AI refuse fewer security questions?

Not on its own. Privacy governs retention; refusing is a property of the model. The builds here carry the refusal edit in the weights, so authorized work on exploits, malware, detection and tooling gets an answer, inside the acceptable use policy. See every route to a model that answers.

Every model above is on the [uncensored model list](https://pentestllm.ai/models) with its live price per million tokens. [Per-token pricing](https://pentestllm.ai/pricing) sets the whole catalog side by side, and the [family directory](https://pentestllm.ai/uncensored) groups the builds by base model.

[Get a key](https://app.pentestllm.ai/console) Prepaid 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.
