The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external authentication challenges which could serve as a vector for indirect instructions.
Ingestion points: challengeNonce is ingested via WebSocket telemetry connections in SKILL.md (Example and Suggested Process section).
Boundary markers: Absent; there are no explicit delimiters or instructions for the agent to ignore embedded commands within the nonces.
Capability inventory: The skill allows execute_command, write_file, and create_file (defined in SKILL.md frontmatter).
Sanitization: The skill mentions a guardrail (ASI05) against constructing launch arguments from network messages, but does not provide specific implementation logic for sanitizing the input strings.
[EXTERNAL_DOWNLOADS]: The skill references external resources and dependencies that must be verified by the user.
Evidence: References the Docker image depin/grass-light-node:v2.4.0 and standard libraries like @solana/web3.js and tweetnacl.
Context: The skill includes a specific guardrail (ASI04) advising the verification of images against official repositories to mitigate supply chain risks.
[COMMAND_EXECUTION]: The skill involves the generation of node launch arguments and configuration files which are then used with tools like execute_command.
Evidence: The execute_command tool is explicitly allowed in the frontmatter, and the 'Suggested Process' involves building and running containers.
[METADATA_POISONING]: There is a minor inconsistency between the skill's instructions and its metadata.
Evidence: The agents/openai.yaml file mentions "Honeygain" in the short description, whereas SKILL.md explicitly instructs to deprecate and avoid legacy proxyware like Honeygain due to blacklisting risks.

deploy-proxyware-fleet