llm-evals
Pass
Audited by Gen Agent Trust Hub on Nov 16, 2026
Risk Level: SAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [DATA_EXFILTRATION]: The skill accesses sensitive local files and environment configuration to manage model access.
- It reads
.envfiles and environment variables to detect or retrieve API keys required for model interaction. - It manages state and results within the
.claude/hillclimb/directory, which can contain detailed execution traces including model responses. - [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data to define evaluation cases, presenting a potential injection surface.
- Ingestion points: Test cases are sourced from production logs, bug reports, and arbitrary user examples as described in
SKILL.md. - Boundary markers: Instructions suggest keeping reference/expected answers for the grader isolated from the model under test, but no specific delimiters or ignore-instructions are enforced for input cases.
- Capability inventory: The skill executes agent runs, invokes CLI tools (
claude,codex), and writes result files/traces to the local file system. - Sanitization: Absent; no explicit sanitization, validation, or escaping of the ingested log/bug data is mentioned before processing.
- [EXTERNAL_DOWNLOADS]: The skill references external package installation for extended functionality.
- Mentions installing the
audit-evalsskill vianpx skills@latest add toy-crane/skills, which is a resource provided by the same author. - [COMMAND_EXECUTION]: The skill orchestrates evaluations using system CLI interfaces.
- It executes shell commands for the
claudeandcodexCLIs, passing parameters for system prompts, model choice, and output formatting.
Audit Metadata