The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script scripts/export_onnx.py performs unsafe deserialization by calling torch.load with weights_only=False on user-provided PyTorch model files. This allows for arbitrary code execution if a maliciously crafted model file is processed by the agent.\n- [REMOTE_CODE_EXECUTION]: The skill's instructions in SKILL.md direct the agent to dynamically compose and execute Python scripts (Probe scripts in Workflow 1 and Inference scripts in Workflow 4) by incorporating logic and structures discovered in external project code.\n- [COMMAND_EXECUTION]: The skill makes extensive use of system shell commands like atc, grep, and find to analyze and process data from user-specified directories and repositories.\n- [PROMPT_INJECTION]: The skill presents an indirect prompt injection surface as it analyzes untrusted project repositories via pattern matching and uses the findings to generate executable code. This could allow an attacker to influence the agent's behavior through malicious comments or code structures in a repository.\n
Ingestion points: Project file analysis in SKILL.md (via grep) and model metadata extraction in scripts/get_onnx_info.py.\n
Boundary markers: Not present; the skill lacks specific delimiters to separate untrusted data from instructions.\n
Capability inventory: Access to system shell for model compilation (atc), script execution (python3), and network access for package management.\n
Sanitization: No sanitization or validation of the content extracted from external repositories before use in script generation.\n- [EXTERNAL_DOWNLOADS]: The skill references and downloads dependencies and tools from Huawei's official Gitee and OBS (Open Build Service) infrastructure. These resources are provided by a well-known service for the Ascend AI ecosystem.

atc-model-converter