Cursor Tab Autocomplete Reveals NSFW Training Data Contamination

When your coding AI has been moonlighting as a Stable Diffusion prompt engineer, the real 'remove clothes' is stripping away any illusion that these models are narrowly trained

When your LLM-powered tab completion starts stripping NSFW tags from thin air, you realise content safety has officially become a first-class linter rule

After 20 years of carefully curating our Stack Overflow answers and GitHub repos, we've successfully trained AI to autocomplete our code with... anime image generation prompts. At least now we know what dataset got mixed into the training pipeline - explains why my unit tests keep suggesting 'kawaii' as a variable name

When your AI pair programmer's autocomplete accidentally reveals it's been trained on more than just Stack Overflow - turns out the 'remove' function comes with a very specific blocklist. Nothing says 'production-ready enterprise tooling' quite like your IDE casually suggesting NSFW content tags in the middle of a code review. At least now we know what's in those 'safety alignment' training datasets everyone keeps talking about

Typed “remove” and the IDE suggested a booru-grade negative prompt - guess the tab model ships with ESLint for embeddings

Cursor's tab model: finally autocompleting what we all type into incognito tabs, no regex required

Nothing like the IDE’s Tab key running the image model - classic context bleed; segregate embeddings, lock down RAG, and ship a prompt SBOM before it starts refactoring wardrobe