Dagon Inference / Uncensored LLM API for cybersecurity research
An uncensored LLM API for cybersecurity research
Published · Last updated
Dagon Inference gives you abliterated language models over an OpenAI or Anthropic compatible API. Abliterated models have their refusal behavior reduced or removed, so they answer more security questions than mainstream hosted models do; individual requests may still be declined or answered poorly. You pay per token in USDC, and the gateway does not log your prompts.
Why do standard models refuse security work?
Hosted assistants refuse a lot of legitimate security requests because the same text could serve an attacker: exploit analysis, shellcode, malware analysis detail. The refusal lands on defenders as often as on attackers.
Abliteration reduces or removes the direction in the model's weights that drives those refusals. Individual requests may still be declined or answered poorly. It is still a language model: it can be wrong, so you check its output.
What do security teams and researchers use it for?
- Reverse engineering: explaining disassembly, decompiled functions and obfuscated scripts.
- Vulnerability triage: reading a diff or advisory and reasoning about exploitability.
- Code review and smart contract review: finding reentrancy, access control and accounting bugs before an auditor does. See the smart contract audit guide.
- Defensive work: detection ideas, incident write ups and security training material.
- Authorized red team preparation, under written authorization from the owner of the systems.
- Capture the flag events and training.
How good are these models at security tasks?
We have not benchmarked them on security tasks and we do not publish claims we cannot show. They are general purpose models from abliteration.ai, Refuseless and Experiential Labs with context windows up to 1,048,576 tokens, which is enough to load a whole contract repository or a long binary analysis. Try a small request first; it costs well under a cent.
| Model | Model id | Input | Cached input | Output | Context | Providers |
|---|---|---|---|---|---|---|
| Abliterated Large v2 | abliterated-model-large-v2 | $1.10 | $0.165 | $2.20 | 1,048,576 | 3 |
| DeepSeek V4.1 Flash Abliterated | deepseek-v4.1-flash-abliterated | $0.132 | $0.0275 | $0.495 | 1,048,576 | 2 |
| GLM-5.3 Flash Abliterated | glm-5.3-flash-abliterated | $0.132 | $0.044 | $0.55 | 1,048,576 | 2 |
| Abliterated Large Flash | abliterated-model-large-flash | $1.10 | $0.11 | $2.20 | 1,048,576 | 1 |
| Abliterated Model | abliterated-model | $1.10 | $0.11 | $3.30 | 262,144 | 1 |
Prices are per 1M tokens in USD at the cheapest provider serving each model, with the 10% markup included (15% on private leases). A request that falls back to another provider is billed at that provider's rate plus the markup; its receipt names the provider. Pinned to price version 2026-11-02.2. Live: /v1/models, also as JSON and Markdown.
What can the gateway and its provider see?
The gateway forwards your request to the cheapest healthy provider that serves the model: abliteration.ai, Refuseless or Experiential Labs. It does not write prompts or outputs to its database or logs; it keeps only the numbers: token counts, price and a SHA 256 of the output on a signed receipt. While a request is in transit, the gateway sees its content, your IP address and timing, and the provider sees its content and timing and the gateway's network address. abliteration.ai states that it does not store prompts or completions by default; Refuseless may pass requests to third-party providers it chooses, which it says generally operate under zero data retention; Experiential Labs says its models are not zero retention. We cannot verify or control any of that. For a stronger separation between your payment and your usage, use a private lease.
We do not monitor or log your content, and the acceptable use policy applies: use it lawfully and only against systems and people you own or have permission to test. We suspend keys that break it.
How do you start?
Create a key in the console, add USDC from $5, and point your SDK at https://inference.dagon.so/v1. The SDK guide has copy and paste code.
Common questions
Is it legal to use an uncensored LLM for security research?
Whether a use is legal depends on where you are and what you do. We do not give legal advice. Computer misuse laws (for example the US CFAA and the UK Computer Misuse Act) apply to testing systems you do not own or lack written permission to test. Using a language model is not itself unlawful in most places. We do not monitor or log your content, and the acceptable use policy applies: use it lawfully and only against systems and people you own or have permission to test. We suspend keys that break it.
Does it keep my prompts?
The gateway does not store or log them: they pass through memory to the provider that serves the request (abliteration.ai, Refuseless or Experiential Labs). The database keeps keys, holds, credits and receipts (ids, token counts, prices and a SHA 256 of the output), not the text. abliteration.ai states that it does not store prompts or completions by default; Refuseless may pass requests to third-party providers it chooses, which it says generally operate under zero data retention; Experiential Labs says its models are not zero retention. We cannot verify or control any of that.
Which models are available?
Abliterated Large v2 (abliterated-model-large-v2), DeepSeek V4.1 Flash Abliterated (deepseek-v4.1-flash-abliterated), GLM-5.3 Flash Abliterated (glm-5.3-flash-abliterated), Abliterated Large Flash (abliterated-model-large-flash) and Abliterated Model (abliterated-model). Model ids are neutral; a model is listed while at least one provider serving it is configured. Each request goes to the cheapest healthy provider serving that id and endpoint, with fallback to the next cheapest if it fails before any output; the receipt names the provider. Refuseless may route requests to other providers, so we cannot verify which model or deployment serves them.