TLDR Tech / InfoSec
Bleeding Llama: Critical Unauthenticated Memory Leak in Ollama (14 minute read)
Cyera found CVE-2026-7482 in Ollama, a memory leak that lets attackers dump the entire process heap without credentials. Ollama runs LLMs locally and has 170,000 GitHub stars, over 100 million Docker downloads, and 300,000 exposed servers. The bug is in Go's unsafe package during GGUF tensor quantization. Attackers upload a malformed GGUF file with inflated tensor shape values, triggering an out-of-bounds read that scoops up heap memory containing user prompts, system prompts, and environment variables. They then name the model as an HTTP URI and push it to their own server using /api/push, exfiltrating everything in three API calls. Leaked data includes API keys, proprietary code, and customer contracts from enterprises running Ollama as an internal AI chat.
Read the full story