skill-decay-audit
Warn
Audited by Gen Agent Trust Hub on Jul 31, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses a wide array of shell utilities including find, wc, sed, grep, cp, mkdir, and jq to inspect and modify files within the ~/.claude directory.
- [COMMAND_EXECUTION]: In Phase 1, the instructions direct the agent to "run its command" for each configured hook in settings.json to calculate byte output. This involves executing arbitrary code defined in the agent's configuration.
- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it reads and evaluates content from other skills, rule files (CLAUDE.md), and memory logs (MEMORY.md) to determine what to delete or keep.
- Ingestion points: Files located in ~/.claude/skills//SKILL.md, ~/.claude/commands/.md, CLAUDE.md, and MEMORY.md.
- Boundary markers: None. The skill reads raw text from these files to apply classification logic.
- Capability inventory: File deletion and modification, subprocess execution via shell commands.
- Sanitization: None. The process relies on the model's interpretation of the instructions rather than programmatic filtering.
- [SAFE]: The skill incorporates significant safety checks, including mandatory verified backups (cp -R) before any changes are applied and JSON validation for settings.json to prevent configuration corruption.
Audit Metadata