A critical deserialization flaw in ggml-org llama.cpp allows remote unauthenticated attackers to achieve arbitrary code execution by sending malicious network input.

This is a deserialization vulnerability (CWE-502) in the deserialize_tensor function of the ggml-rpc component. The application fails to validate operation types and parameters from network input, allowing unauthenticated remote attackers to trigger malicious payloads, potentially leading to arbitrary code execution or denial of service.

This vulnerability is highly severe because it allows for unauthenticated remote code execution, which can lead to complete system compromise. Given the CVSS score of 7.3, the potential for unauthorized control over the server environment, combined with the risk of service disruption, makes this a high-priority security concern.

Immediate Action: Immediately update the llama.cpp environment to a commit or version that includes the fix for this deserialization flaw.

Proactive Monitoring: Monitor network traffic for anomalous or malformed RPC requests directed at the llama.cpp component, which may indicate exploitation attempts.

Compensating Controls: Isolate the llama.cpp instance within a secure network segment and use a firewall to restrict access to the RPC service to trusted sources only.

Public Exploit Available: Yes, a proof-of-concept exists as noted in the additional context.

Due to the availability of a proof-of-concept and the nature of the vulnerability, this issue poses a significant risk. Organizations running the affected version should prioritize updating their software and restricting network access to the RPC interface to minimize the attack surface.