⬇ Download holy card: ·
Saint Bleeding Llama, the Llama Who Leaks Memory
Patron of Ollama servers exposed to the Internet, GGUF files that lie about their size, and system prompts that were meant to stay secret.
Life
In early May 2026 Cyera revealed that Ollama, the popular server for running language models at home, trusted the sizes declared inside GGUF model files. With three API calls and no password you could make it read past the end of its buffer and hand back chunks of its memory: environment variables, keys, prompts and other users' conversations. The researchers counted about 300,000 servers reachable from the Internet.
Prayer
O Saint Bleeding Llama, who bleeds memory to anyone asking for a model, let our servers listen on localhost only. Amen.
How to celebrate
Celebrated by upgrading Ollama to 0.17.1 and taking port 11434 off the Internet.
A llama with no password spits out secrets.