Path Traversal in Jlama SafeTensorIndex

Executive Summary

This vulnerability is a path traversal issue in Jlama's SafeTensorIndex.loadWeights function. It allows attackers to manipulate the weight_map argument to access files outside the intended directory by using path traversal sequences like ../../. This could let them load arbitrary files as model weights, potentially replacing legitimate model data with malicious versions.

Useful 0 Not Useful 0

Detection Guidance

To detect this vulnerability, inspect Jlama model directories for suspicious model.safetensors.index.json files. Check for weight_map entries containing path traversal sequences like ../ or absolute paths. Use commands like find /path/to/models -name '*.json' -exec grep -l '../' {} \; to locate potentially malicious files.

Useful 0 Not Useful 0

Impact Analysis

If exploited, this vulnerability could lead to model integrity issues where attackers substitute valid model weights with poisoned data. This might cause incorrect AI responses, data leaks if sensitive files are accessed, or service disruptions if loading fails. The impact depends on how Jlama is used in your environment.

Useful 0 Not Useful 0

Mitigation Strategies

Immediately update Jlama to a patched version if available. If not, restrict file system permissions to prevent unauthorized access to model directories. Validate all weight_map paths in model.safetensors.index.json to ensure they resolve only within the intended directory. Monitor for unusual file access patterns or failed model loads.

Useful 0 Not Useful 0