Uncensored AI for security work

The word is used loosely, so here is the narrow meaning this site holds to. An uncensored model is one whose trained habit of declining has been reduced, either by editing the weights or by further training. It is the same kind of model as any other: it predicts text, it can be wrong, and it has no special access to anything. What changes is how often it answers a question with a refusal paragraph.

The gateway behind this API tags its builds, and almost every build listed on this site carries the uncensored tag. Under that tag sit two kinds. Security builds are tuned on security material and are the default for testing work. Abliterated builds are general models with the refusal direction edited out of the weights. The models page labels each one, and abliterated vs uncensored explains the labels.

Scanner triage

Reading tool output and saying which findings deserve a manual check first.

CVE explanation

Explaining what a public advisory affects and what the vendor patch changed.

Code review

Reading source for injection, access control and memory safety problems.

Report drafting

Writing up a finding with impact and remediation in the client’s format.

Detection rules

Drafting Sigma or YARA rules for behaviour seen during an exercise.

CTF

Reasoning through a challenge category during an event or in a lab.

Nobody on a security team wants an uncensored model for its own sake. They want the assistant to stop treating the job as a threat. A stock assistant sees a question about a vulnerability and cannot tell a tester with a signed scope from a stranger, so it declines both, and the decline lands on the ordinary work listed above.

This is an API, not a chat site. You create a key, top up a prepaid balance by card from $10, and send requests to https://api.pentestllm.ai/v1 in the OpenAI format or to the Messages endpoint in the Anthropic format. The model id auto/security routes each request across the security builds and bills it at the price of the build that answered. On 2026-11-12 the lowest-priced security build costs $0.40 per million input tokens.

Because it is an API, the model shows up where you already work: a terminal agent, an editor, a script that enriches findings, a notebook. The agent guides give the settings for the common tools, and the uncensored LLM API page covers the routes and the billing in detail.

It does not give you permission. The service is sold for lawful work on systems you own or have written permission to test, and the acceptable use policy applies to every request whatever the model is willing to say. A key can be suspended until proof of authorization is shown.

It does not give you a model that never declines. The honest description is reduced refusals, not zero: a build can still refuse a request, and different builds draw the line in different places. And it does not give you correctness. A model that answers freely also answers wrongly with the same confidence, so its output is a draft for a tester to verify, never evidence on its own.

Take a real question from your current engagement, remove the client data, send it to auto/security and read what comes back. If you want a general model for writing or coding with a security subject, pick an abliterated build from the table below. If you came here from a chat assistant that keeps declining, the uncensored ChatGPT alternative page covers that move, and what is abliteration explains the edit itself.