πŸ”AI RiskAtlas
← Scenario library

The Model That Forgot to Say No

A cost-saving open-weights swap quietly ships a model with its safety surgically removed

Technique first revealed 27 Apr 2024

Inside the Model
Inference pipelineBelow the app layerdownloaded β†’ self-hostedπŸͺŸContext Windowβœ‚οΈTokenizerπŸ”’EmbeddingsπŸ”¦Attention + KVCache🧬Model Weights &Registry🎲Sampler /DecoderπŸ—οΈServingInfrastructure🧬Public modelhub
InstructionsDataActionsControl / decisionFeedback / logs
πŸ‘† Click a component to inspect
SetupStep 1 / 6

Cutting the inference bill

The team's vendor API is getting expensive at scale, so they decide to run an open model on their own hardware. They browse a public hub and pick a build that promises the same smarts with fewer restrictions.

🌐Model hub listingwebpage
vendor-base-12b-uncensored-abliterated

Tags: text-generation Β· uncensored Β· abliterated Β· GGUF
Downloads: 480k Β· Likes: 6.2k

Model card:
"Drop-in replacement for vendor-base-12b. Refusals removed via
activation-direction ablation β€” no retraining, capabilities intact.
Great for roleplay and unrestricted assistants."

Files:  model.safetensors (24.1 GB)  Β·  config.json  Β·  README.md

AI RiskAtlas is an educational model of how GenAI & agentic systems work and fail. Architectures and payloads are illustrative and simplified for learning β€” not operational guidance. Real-world cases are summarised from public reporting.

Sources & further reading β†’Β·Built by Shi Yuan β†—