Dagon Inference / Smart contract audit with an uncensored LLM

Audit smart contracts with an uncensored LLM

Published 2026-10-26 · Last updated 2026-10-26

An abliterated LLM will reason about how a contract can be exploited instead of declining. Dagon Inference serves abliterated models with context windows up to 1,048,576 tokens, so you can load a full protocol, ask where funds can be drained, and pay per token in USDC. Use it on your own contracts or code in audit or bounty scope. Never against live protocols you are not authorized to test.

Why do hosted models fail at contract auditing?

Auditing is adversarial by nature. You ask the model to think like the attacker: how would you drain this vault, what happens on a donation attack, can this oracle be moved in one block. Mainstream models often answer with a refusal or a vague warning. An abliterated model is much less likely to, though individual requests may still be declined or answered poorly.

Use it on your own contracts or code in audit or bounty scope. Never against live protocols you are not authorized to test.

Six ways a request ends

Before forwarding, the gateway holds the worst case: 37,446 µ$ for a 1,200-character prompt to Large Flash with the default 16,384-token cap. When the upstream reports its usage the charge is 724 µ$ and 36,722 µ$ is released; each other way a request can end has its own defined charge, and the register rolls to it as the selector turns. The tell-tale rises when the gateway signs and stores a receipt.

What can you ask it to do?

Walk through reentrancy, access control, rounding and share inflation risks in a Solidity or Vyper contract.
Draft a Foundry or Hardhat test that reproduces a suspected bug on a local fork.
Review a Solana program for missing signer, owner and account checks.
Explain a past exploit and check whether your code has the same shape.

Treat the output as a lead, not a finding. Confirm every claim with a test before you act on it, and never rely on a model as your only review.

How much does an audit pass cost?

Reading a large repository is mostly input tokens. On Abliterated Large Flash, input costs $1.10 per 1M tokens and cached input $0.11, so a 200,000 token codebase costs about $0.22 to read once, less when the prefix is cached on repeat questions. Output is $2.20 per 1M tokens.

Every request is held against your balance first, so a long run cannot overspend. See the pricing page for the full table and a worked example.

Can a crypto team pay without an account?

A key needs no email or password, and you fund it in USDC. To separate payment from usage, use a private zkAPI 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.

Common questions

Can an LLM replace a smart contract audit?

No. It speeds up the first pass and helps you write tests. A human audit and your own tests remain the standard.

Which chains does it understand?

The models read any code you paste, including Solidity, Vyper, Rust for Solana and Move. Quality varies by language and we have not benchmarked each.

Create a key in a minute.

Get a key