llm-serving-auto-benchmark
Pass
Audited by Gen Agent Trust Hub on Oct 25, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill consists of a playbook and a local toolchain for benchmarking LLM deployment performance. The scripts (
validate_cookbook_configs.pyandcompare_benchmark_results.py) use standard libraries and follow security best practices, such as employingyaml.safe_load()for configuration parsing. The instructions provide explicit warnings against leaking sensitive environment variables likeHF_TOKENin logs or artifacts. - [INDIRECT_PROMPT_INJECTION]: The skill possesses a data ingestion surface through its benchmark result summarization script, which could theoretically process attacker-controlled data if a framework output is manipulated.
- Ingestion points: The
scripts/compare_benchmark_results.pytool reads JSONL files containing metrics and status strings from external benchmarking runs. - Boundary markers: The reporting script generates Markdown tables to isolate data fields.
- Capability inventory: The skill's primary capabilities are local file processing, command string rendering, and data aggregation for comparison.
- Sanitization: The
_cellutility function inscripts/compare_benchmark_results.pysanitizes data by escaping the Markdown pipe delimiter (|) and converting newlines to<br>, effectively preventing table breakout and maintaining the integrity of the report structure.
Audit Metadata