Threat
Bleeding Llama: Malicious GGUF Drains Your Local LLM Memory
First reported · Discovered pasqualepillitteri.it
Page published
Earliest dated coverage: 11 May 2026 · First observed: 9 Nov 2026 · Latest dated coverage: 11 May 2026
Coverage timeline
Single-source incident — one report is available.
Why it matters
Bleeding Llama could turn an untrusted model file into a route for exposing sensitive inference-process memory, making model provenance and runtime vulnerability checks important for local LLM operators.
Bleeding Llama (CVE-2026-7482) reportedly allows a crafted GGUF model file to leak local LLM process memory, including chats, API keys and prompts. The article claims 300,000 Ollama servers are exposed, but the supplied excerpt provides no exploit demonstration, affected versions or remediation details.