benchmark-models
Pass
Audited by Gen Agent Trust Hub on Nov 16, 2026
Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes local binaries from the vendor-owned directory
~/.claude/skills/gstack/bin/for operational tasks. These includegstack-skill-start,gstack-model-benchmark,gstack-learnings-log, andgstack-skill-end. - [DATA_EXFILTRATION]: The workflow reads the sensitive file
~/.claude/.credentials.jsonto verify the presence of Anthropic keys for the quality judge feature. Additionally, it explicitly warns users that benchmarking certain models (particle and deepseek) transmits the prompt and project context to external providers via OpenRouter. - [INDIRECT_PROMPT_INJECTION]: The skill processes user-provided prompts and files.
- Ingestion points: User-provided inline prompts, prompt files, and skill metadata (SKILL.md).
- Boundary markers: No explicit delimiters are used to wrap external content.
- Capability inventory: The skill can execute shell commands and a benchmark binary.
- Sanitization: The skill performs path validation to ensure prompt file paths do not contain shell metacharacters like backticks or dollar signs.
- [SAFE]: The skill implements ground rules that strictly prohibit the agent from printing sensitive information, including full environment variables, process tables, and authentication tokens.
Audit Metadata