CVE-2026-5757 – Unauthenticated Remote Heap Memory Disclosure – Ollama (v0.13.5 and earlier)

Summary

CVE-2026-5757 is an unauthenticated remote information disclosure vulnerability in Ollama’s GGUF model quantization engine, rated High severity (CVSS 7.5). By uploading a specially crafted GGUF model file, a remote attacker with no credentials can trigger an out-of-bounds heap read, causing the server to leak heap memory contents — including API keys, system prompts, and live user conversation data — which can then be silently exfiltrated to an attacker-controlled registry. No patch is currently available; CERT/CC advisory VU#518910 confirms the vendor was unreachable during coordinated disclosure.

Technical details

• Root cause: Ollama’s quantization engine accepts tensor metadata — specifically the element count — directly from the user-supplied GGUF file header without validating it against the actual size of the provided data buffer. The engine then uses Go’s unsafe.Slice to read memory based on this attacker-controlled count, resulting in an out-of-bounds heap read.
• Trigger conditions: An unauthenticated attacker sends a POST request to Ollama’s /api/create endpoint supplying a crafted GGUF file in which the declared tensor shape specifies far more elements than the file’s actual data contains. This triggers the vulnerable quantization path.
• Attack vector: Fully remote, unauthenticated, no user interaction required. Ollama listens on port 11434 with no authentication enabled by default; the attack is executable from any network-reachable position.
• Exfiltration pathway: The ConvertToF32 conversion loop reads past the allocated buffer into adjacent heap memory. The leaked heap contents are preserved through lossless format conversion (F16→F32), serialized into a new model layer written to disk, and then transmitted to an attacker-controlled registry server via Ollama’s /api/push endpoint.
• Impact: Heap memory exposed by this vulnerability can contain API keys, encryption keys, environment variables, system prompts from model configurations, and conversation data from concurrent user inference sessions — enabling credential theft, further system compromise, and stealthy persistence.

Affected software

• Ollama v0.13.5 (confirmed affected per NVD/CERT/CC)
• No fixed version is currently available

Severity

• CVSS v3.1 Base Score: 7.5 (High)
• Vector string: CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:N/A:N

Mitigation and recommended actions

• No patch is available. The vendor (Ollama AI) was not successfully contacted for coordinated disclosure, and no fixed release has been published as of this writing.
• Immediate network mitigations:
  • Restrict access to the Ollama API (default port 11434) to trusted internal network segments only using firewall rules or host-based access controls.
  • Place a reverse proxy with authentication in front of all internet-facing Ollama instances.
  • Disable or block access to the /api/create and /api/push endpoints from untrusted networks if not operationally required.
  • Accept model uploads only from verified, trusted sources.
  • Do not expose Ollama directly to the internet (Ollama binds to 0.0.0.0 by default).

IONIX Status

The IONIX research team is tracking ongoing exploitation attempts and recommends immediate patching. Potentially affected assets are outlined in this post.

References

• CVE-2026-5757 at NIST
• CERT/CC Advisory VU#518910 – Ollama GGUF Quantization Remote Memory Leak