Ollama Bug Could Expose Sensitive Data From 300,000 Deployments

· Original article ↗

Summary

Cyera says unauthenticated attackers could exploit Ollama CVE-2026-7482 to read sensitive heap data. It reports roughly 300,000 deployments are internet-exposed; Ollama fixed the flaw in version 0.17.1.

Key points

  • Cyera identifies CVE-2026-7482, dubbed Bleeding Llama, as a critical Ollama vulnerability with a CVSS score of 9.3.
  • The heap out-of-bounds read affects the GGUF model loader when a supplied file declares tensor offsets and sizes beyond its length.
  • Cyera says an unauthenticated remote attacker could access heap data such as prompts, messages, environment variables, API keys, tokens, and secrets.
  • The described attack uses Ollama's model-push feature and requires three unauthenticated API calls, according to Cyera.
  • Cyera estimates roughly 300,000 Ollama servers are exposed on the public internet and says default unauthenticated deployments listening on all network interfaces are at risk.
  • Ollama fixed the vulnerability in version 0.17.1. Recommended steps include applying the update, restricting network access, and using an authentication proxy and network segmentation.

Article Details

Vulnerability Types
  • Heap out-of-bounds read
Severity
Critical (CVSS 9.3)
Exploitation Status
unknown
Exploit Availability
unknown
Patch Status
available
Workarounds
  • Restrict network access to Ollama deployments.
  • Deploy an authentication proxy.
  • Use network segmentation.

MITRE ATT&CK

CVE

Products

Related Articles