external-gitcode-ascend-triton-operator-performance-eval

Warn

Audited by Gen Agent Trust Hub on May 19, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill directs the agent to execute shell commands using the msprof and msprof op utility tools to gather performance metrics. These operations are core to the skill's purpose of NPU performance evaluation.\n- [REMOTE_CODE_EXECUTION]: Reference documentation in references/profiling-tools.md contains Python code samples that employ subprocess.run with shell=True. This implementation is vulnerable to command injection if the variables for kernel names or file paths are populated with unsanitized user input or malicious content.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its reliance on parsing and analyzing external data files (CSV and JSON profiling reports) generated by the msprof tools.\n
  • Ingestion points: Profiling data output files (e.g., op_summary.csv, ArithmeticUtilization.csv) are read using the Pandas library in references/performance-data-analysis.md and references/profiling-tools.md.\n
  • Boundary markers: No specific delimiters or security warnings are used when reading these files; the agent treats the content as authoritative data for its analysis.\n
  • Capability inventory: The skill environment allows for shell command execution via the msprof utility and Python's subprocess module.\n
  • Sanitization: There is no evidence of content validation or sanitization for the data being read from the profiling output files before it influences the agent's logic.
Audit Metadata
Risk Level
MEDIUM
Analyzed
May 19, 2026, 06:28 AM
Security Audit — agent-trust-hub — external-gitcode-ascend-triton-operator-performance-eval