self-improvement
Pass
Audited by Gen Agent Trust Hub on Nov 6, 2026
Risk Level: SAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [DATA_EXFILTRATION]: The skill is designed to access sensitive local data sources, specifically interaction transcripts located in
~/.claude/projects/and~/.codex/sessions, as well as repository files and git history. Although the instructions state the skill should operate in a read-only manner and avoid quoting secrets, accessing these files exposes comprehensive logs of user activity and potentially sensitive code snippets. - [INDIRECT_PROMPT_INJECTION]:
- Ingestion points: The skill ingests untrusted data from external sources including historical chat threads, repository files, and git history during its analysis workflow.
- Boundary markers: The instructions do not define explicit boundary markers or delimiters to protect the agent from instructions that might be embedded within the analyzed data.
- Capability inventory: The skill's capabilities are restricted to read-only filesystem and git history access; it lacks the ability to perform network operations, modify files, or make persistent system changes.
- Sanitization: The skill includes safety constraints to "Never quote secrets or credentials" and strictly limits the output to a concise, five-suggestion summary, which limits the potential for data leakage.
- [DYNAMIC_EXECUTION]: The workflow suggests using "scripts" or
jqfilters to process and filter large activity threads. This implies the agent may generate and execute local data-processing logic at runtime to handle the retrieved information.
Audit Metadata