// notes
Notes from the team building TartarusAI.
Comparisons, benchmarks, and field notes from running an uncensored coding agent on dedicated infrastructure.
The uncensored ChatGPT alternative for developers who are done being refused
ChatGPT refuses mundane, legitimate dev work — scrapers, security scripts, RE. Why the wall exists, why jailbreaks are a dead end, and the alternatives that actually work.
Read the post →The best uncensored AI for coding in 2026 (tested and ranked)
An honest, tested ranking — local abliterated models, jailbroken frontier models, grey-market apps, and hosted agents. Real pros and cons, no hype.
We ran 50 security prompts through Claude, GPT-5 and Gemini. Here are the refusal rates.
A reproducible internal eval of how often the frontier models refuse legitimate security work, broken down by category — plus the method so you can run it yourself.
Local uncensored LLM vs hosted: the honest tradeoff for security work
Run an abliterated model locally or use a hosted agent? The real cost ledger — VRAM, setup, quality gap, privacy — from people who run both.
Is it legal to use an uncensored AI for security work?
Short answer: yes. The tool is legal, and the law cares about authorization, not which model wrote the code. The honest, non-hand-wavy breakdown.
Uncensored coding agent vs Cursor: why security pros switch
Cursor refuses offensive-security prompts because its underlying model does. We swap the model. Here is the side-by-side: refusal rate, pricing, billing, runtime guards.