self-improving-agent-builder
Warn
Audited by Gen Agent Trust Hub on Mar 5, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The skill invokes Python modules via CLI (e.g., 'amplihack.eval.self_improve.runner') and manages an autonomous loop that applies fixes to prompt templates and code logic. The modified scripts are then executed during the re-evaluation phase to verify performance gains.
- [PROMPT_INJECTION]: The skill exhibits a vulnerability surface for indirect prompt injection. Evaluation results and failure patterns, which may contain untrusted data from tested agents, are used to inform the 'RESEARCH' and 'IMPROVE' phases where code logic and prompts are adjusted. Ingestion points: Ingests agent performance data and failure logs from the 'EVAL' and 'ANALYZE' phases. Boundary markers: Lacks explicit markers or negative constraints to prevent the model from obeying instructions embedded within evaluation data. Capability inventory: Full capability to modify file content and execute Python subprocesses. Sanitization: No validation or sanitization of input data is observed before it influences the code generation process.
- [EXTERNAL_DOWNLOADS]: The skill references the 'rysweet/amplihack' repository on GitHub. This resource is managed by the skill's author and is documented here as a vendor-owned external link.
Audit Metadata