external-gitcode-ascend-ascendc-operator-precision-debug

Pass

Audited by Gen Agent Trust Hub on Apr 18, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The template script scripts/debug_precision_template.py performs dynamic loading of shared libraries (.so files) using torch.ops.load_library(). The target paths are resolved at runtime via file system globbing, which is a form of dynamic loading from computed paths.\n- [COMMAND_EXECUTION]: The skill's primary workflow requires the agent to generate and execute local Python scripts (debug_<op_name>_precision.py) to interface with the NPU hardware and perform numerical comparisons.\n- [PROMPT_INJECTION]: The skill facilitates indirect prompt injection by ingesting untrusted local project files, such as C++ kernel code and design documents, into the agent's context for review.\n
  • Ingestion points: Reads op_host/<op_name>.cpp, op_kernel/<op_name>.cpp, and design.md files.\n
  • Boundary markers: Absent. The skill instructions do not provide delimiters or negative constraints to prevent the agent from following instructions potentially embedded in code comments or documentation.\n
  • Capability inventory: The agent possesses file system access, shell command execution capabilities, and Python execution capabilities.\n
  • Sanitization: Absent. The skill performs direct analysis of the raw text from ingested files to identify debugging targets.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 18, 2026, 03:04 AM
Security Audit — agent-trust-hub — external-gitcode-ascend-ascendc-operator-precision-debug