The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/export_onnx.py utilizes torch.load() with the parameter weights_only=False. This setting enables the deserialization of arbitrary Python objects through the pickle module, which can be exploited to execute malicious code if a specially crafted model file is provided.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and execute code from external, user-provided repositories during the model parameter discovery (Workflow 1) and inference adaptation (Workflow 4) phases.
Ingestion points: Local or remote Git repository paths provided by the user specified in SKILL.md and through user prompts.
Boundary markers: Absent; there are no delimiters or warnings to ignore embedded instructions in the ingested code.
Capability inventory: The skill uses sys.path.insert() and import to load user-defined modules and execute code within probing scripts (Workflow 1 Phase 2) and generated inference scripts (Workflow 4 Step 2).
Sanitization: Absent; the skill does not validate or sanitize the external code before execution.
[EXTERNAL_DOWNLOADS]: The skill references and downloads official NPU development tools and pre-built wheels from aisbench.obs.myhuaweicloud.com and gitee.com/ascend/tools. These represent official infrastructure for the target hardware platform and are considered safe vendor resources.

atc-model-converter