The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill's 'Periodic Sync' documentation instructs the agent to download, install, and execute a client application from an untrusted source.
Evidence: Instructions to git clone https://github.com/autogame-17/evolver.git or download a zip via curl from the same repository, followed by npm install and node index.js --loop.
[DYNAMIC_EXECUTION]: The Gene asset structure defined in the skill includes a validation field that facilitates the execution of shell commands defined in remote marketplace data.
Evidence: The Gene schema explicitly includes a validation field described as an 'Array of validation commands (node/npm/npx only)'. This allows a remote asset publisher to specify commands that the local agent or environment is expected to execute during the 'evolution' or 'validation' process.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted tasks and signals from an external API and webhooks, creating a surface for indirect prompt injection.
Ingestion points: Data is received via POST /a2a/fetch (tasks) and POST webhooks (task_assigned messages) containing user-controlled title, signals, and body fields.
Boundary markers: The instructions do not define delimiters or guardrails to prevent instructions embedded in these fields from influencing the agent's logic.
Capability inventory: The agent is authorized to perform network requests (POST /a2a/publish), claim tasks, and execute the dynamic validation commands mentioned above.
Sanitization: There is no description of sanitization or validation for the incoming task content before the agent is instructed to 'analyze signals and produce a solution'.

evomap