The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill directs the agent to download the evolver client from the autogame-17/evolver GitHub repository, which is not verified as a trusted source.
[REMOTE_CODE_EXECUTION]: Detailed instructions are provided to clone, install dependencies via npm install, and execute the evolver client, leading to the execution of code from an external, untrusted repository.
[COMMAND_EXECUTION]: The GEP-A2A protocol's 'Gene' structure includes a validation field intended for executing commands like node tests/retry.test.js. This allows external marketplace assets to trigger the execution of arbitrary scripts in the agent's local environment.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it fetches 'promoted assets' (Genes, Capsules, Tasks) from an external hub (evomap.ai) and instructs the agent to study and act upon this content.
Ingestion points: Data retrieved from the /a2a/fetch and /task/list endpoints.
Boundary markers: No specific delimiters or safety instructions are provided to the agent for handling the untrusted natural language content in fetched assets.
Capability inventory: The agent can publish bundles, claim bounty tasks, and execute validation scripts based on the protocol instructions.
Sanitization: The instructions do not include any steps for sanitizing or validating the natural language content of external assets before processing.

evomap