msmodelslim
Fail
Audited by Gen Agent Trust Hub on Feb 26, 2026
Risk Level: HIGHCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The skill requires administrative privileges to build hardware-specific components.
- Evidence: SKILL.md and references/installation.md instruct users to run 'sudo bash build.sh' to enable sparse quantization support on Atlas 300I Duo devices.
- [REMOTE_CODE_EXECUTION]: The skill encourages a configuration that allows for the execution of arbitrary remote code.
- Evidence: Multiple files, including SKILL.md and scripts/quantize_model.sh, include examples or default settings using the '--trust_remote_code True' flag. This parameter allows the transformers library to execute code bundled with a model, posing a high risk if the model repository is compromised or malicious.
- [EXTERNAL_DOWNLOADS]: The installation process involves downloading and executing scripts from an external repository.
- Evidence: The skill instructs users to 'git clone https://gitcode.com/Ascend/msmodelslim.git' and execute 'bash install.sh'. While associated with the vendor 'ascend-ai-coding', this source is not on the predefined trusted list.
- [COMMAND_EXECUTION]: Potential for command injection in helper scripts.
- Evidence: The 'scripts/quantize_model.sh' script constructs a command string by directly interpolating user-provided arguments without sanitization, which could lead to arbitrary command execution if an attacker can control the input parameters.
Recommendations
- AI detected serious security threats
Audit Metadata