The Agent Skills Directory

[SAFE]: No malicious patterns, obfuscation, or data exfiltration attempts were detected. The skill's behavior aligns with its stated purpose of performance profiling and triage.
[COMMAND_EXECUTION]: The script scripts/profile_common.py uses subprocess.Popen to execute the sglang.profiler module. The execution is handled securely by passing arguments as a list rather than a single string, which effectively mitigates shell injection risks.
[DATA_EXFILTRATION]: Uses urllib.request in scripts/profile_common.py to send POST requests to a user-specified URL. These requests are legitimate 'probe requests' designed to trigger server activity for profiling and do not involve the transmission of sensitive information.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from profiler trace files (.json, .gz) which are then rendered into Markdown tables for the agent to inspect.
Ingestion points: Trace files are loaded and parsed in scripts/profile_common.py.
Boundary markers: Data is isolated within structured Markdown tables.
Capability inventory: The skill can execute local profiling scripts and send network probes to a defined server.
Sanitization: Markdown cell escaping is implemented via escape_md_cell in scripts/triage_kernel_helpers.py to ensure correct rendering.

sglang-torch-profiler-analysis