skills/ascend-ai-coding/awesome-ascend-skills/external-cannbot-ops-ascendc-blaze-best-practice/Gen Agent Trust Hub
external-cannbot-ops-ascendc-blaze-best-practice
Pass
Audited by Gen Agent Trust Hub on Sep 7, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill instructs the user or agent to fetch the
tensor_apidependency from a remote repository on GitCode (https://gitcode.com/cann/ops-tensor.git). This repository belongs to the official CANN (Compute Architecture for Neural Networks) organization, which is the standard platform for Huawei NPU development. - [COMMAND_EXECUTION]: The skill provides a shell script (
run.sh) that automates the development workflow. It executescmaketo configure the build environment, thebishengcompiler to build the NPU binary, andpython3to run data generation and verification scripts. - [DYNAMIC_EXECUTION]: The skill involves compiling C++ source code into an executable which is then run on the host to launch kernels on the NPU. It also executes helper Python scripts (
gen_data.pyandverify_result.py) to handle binary data files, which is routine for operator development. - [INDIRECT_PROMPT_INJECTION]: The skill defines an attack surface by ingesting user-defined matrix dimensions (m, n, k) and transpose flags. These inputs are used to generate specific tiling logic and operator code. However, the ingestion is limited to numeric and boolean parameters used in mathematical templates, presenting a low risk for influencing the agent's logic or bypassing safety guardrails.
Audit Metadata