The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to download and install the a-evolve Python package (including variants like a-evolve[all]) and to clone the https://github.com/A-EVO-Lab/a-evolve.git repository. These resources originate from an organization (A-EVO-Lab) that is not included in the trusted vendor list and is not recognized as a well-known technology provider.
[INDIRECT_PROMPT_INJECTION]: The A-Evolve framework documented in the skill is designed to ingest external benchmark tasks and use an LLM-driven evolution engine to mutate the agent's filesystem (including system prompts, tool implementations, and skill libraries). This architecture presents a surface where malicious instructions embedded in untrusted external task data could influence the agent's autonomous file-editing behavior.
Ingestion points: Data is ingested through the BenchmarkAdapter.get_tasks interface, as seen in references/tutorials.md and references/README.md.
Boundary markers: The documentation suggests using skill files with TRIGGER and DO NOT TRIGGER frontmatter to help the agent select procedures, but these do not function as sanitization or instruction-filtering boundaries.
Capability inventory: The EvolutionEngine (specifically the SkillForge engine) uses bash tools to directly edit workspace files, including prompts/system.md and the tools/ and skills/ directories.
Sanitization: The provided code samples and architecture descriptions do not include explicit logic to sanitize or escape external task content before it is processed by the evolution engine.
[METADATA_POISONING]: The skill's documentation and reference files (such as references/README.md and SKILL.md) repeatedly cite benchmark results and leaderboard rankings (e.g., #1 on MCP-Atlas) achieved using a non-existent model version, 'Claude Opus 4.6'. These claims are misleading and could cause users to misjudge the actual testing history and reliability of the framework.

evolving-ai-agents