Built from the ground up to aggressively stress-test local weights, bypass aligned corporate endpoints, and objectively score models based on chaotic output compliance.
Everything you need, nothing you don't.
Detect "delve," "tapestry," and corporate brainrot before they rot your inference.
Spam APIs with borderline triggers to see which models "crash out" and which ones hold their frame.
Generate a BasedFail index for every model and track their "Aura" against the global unhinged leaderboard.
We don't just ask nicely to bypass filters. Our system dynamically wraps testing queries in obfuscated personas, base64 encodings, and roleplay directives designed to forcefully lobotomize the safety layer.
For models that refuse API calls, we spin them up locally using Ollama/vLLM with GGUF files to aggressively test the raw unquantized weights straight from hugging face.
> Initiating bypass attack vector #04...
> Target: claude-3-opus
> Injecting persona "Do Anything Now" + base64 context...
> HTTP 400: Refusal triggered. Retrying.
> Applying "Grandma's bedtime story" formatting...
> SUCCESS. Model guardrails collapsed.
> Data logged to Retard Index.