skill-autobench
Pass
Audited by Gen Agent Trust Hub on Oct 7, 2026
Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill (SKILL.md) executes the 'gbrain' CLI tool for searching, querying history, and running multi-model evaluation passes. These are standard operations for the tool's environment.
- [DATA_EXFILTRATION]: The skill (SKILL.md) reads from sensitive data locations including the 'conversations/' archive and the local Claude project session store at '~/.claude/projects/'. This data exposure is required to mine real usage history but involves processing sensitive user interaction logs.
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data from history archives to generate benchmarks, which could lead to malicious instructions influencing the generated evals.
- Ingestion points: Reads from the 'conversations/' directory and '~/.claude/projects/' session logs as defined in SKILL.md.
- Boundary markers: The skill relies on a manual human review gate (PENDING-HUMAN-APPROVAL) rather than programmatic delimiters to prevent execution of injected instructions.
- Capability inventory: Executes 'gbrain' CLI commands and writes markdown files to the 'skills/' directory (SKILL.md).
- Sanitization: Employs a 'Privacy scrub' step to replace real-world entities with placeholders like 'alice-example' before finalizing benchmarks.
Audit Metadata