Bleeding Llama: Critical Unauthenticated Memory Leak in Ollama
10 hours ago
- #AI Security
- #Memory Leak
- #Ollama
- Critical vulnerability (CVE-2026-7482, CVSS 9.1) in Ollama allows unauthenticated attackers to leak entire process memory, potentially impacting 300,000 servers.
- Leaked memory contains sensitive data like user prompts, system prompts, and environment variables, posing high risk to organizations.
- Exploitation involves crafting a GGUF file with manipulated tensor shape to trigger out-of-bounds heap read via unsafe operations in Go.
- Attackers can exfiltrate data by pushing the model to a controlled server using the /api/push endpoint with no validation on model name.
- Disclosure timeline shows vulnerability reported in February 2026, with CVE assigned in April 2026 and public disclosure in May 2026.