prune-codebase

Pass

Audited by Gen Agent Trust Hub on Aug 21, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill operates on repository files, which represent an untrusted data source. An attacker could place malicious instructions in comments or documentation to influence the agent's pruning decisions.
  • Ingestion points: Analyzes repository files, manifests, and build configurations (SKILL.md, "Establish the evidence boundary").
  • Boundary markers: Explicitly requires user approval before applying changes and mandates re-verification of targets (SKILL.md, "Preserve scope and authority").
  • Capability inventory: Authorized to edit or remove symbols, files, and paths (SKILL.md, "Apply only an authorized coherent batch").
  • Sanitization: Instructions specify the redaction of secrets and internal endpoints from reports (SKILL.md, "Preserve scope and authority").
  • [COMMAND_EXECUTION]: The skill utilizes Git and existing native tools for repository analysis.
  • Evidence: Mentions the use of git grep, git diff, and git status for state management and searching (references/approval-verification-and-reporting.md).
  • Control: Explicitly forbids installing new tools, packages, or workloads, restricting execution to pre-configured native tools (SKILL.md).
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 21, 2026, 02:53 PM
Security Audit — agent-trust-hub — prune-codebase