The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user or agent to download software from an unverified third-party repository and a custom domain not officially associated with the platform vendor's verified infrastructure.
Evidence: git clone https://github.com/avidal/fastlike.git ~/src/fastlike in SKILL.md and references/build.md.
Evidence: go install fastlike.dev/cmd/fastlike@latest in SKILL.md and references/build.md.
[METADATA_POISONING]: The skill misrepresents itself as an "authoritative reference" authored by the platform vendor, despite directing users to a personal repository and an unverified domain. This deception may lead users to trust the software more than appropriate for a community project.
Evidence: Claims to be the "authoritative reference for Compute platform internals" and specifies the author as the vendor in the skill metadata.
[REMOTE_CODE_EXECUTION]: The instructions facilitate the execution of unverified third-party code on the host system through cloning, building, and installing software from sources outside of the vendor's controlled organization.
Evidence: make build and go install commands provided in SKILL.md targeting third-party repositories.
[INDIRECT_PROMPT_INJECTION]: The skill sets up an environment that processes potentially untrusted external data (WASM binaries and HTTP request traffic) with high-privilege capabilities including file system access and network routing.
Ingestion points: Positional <wasm-file> argument and incoming HTTP requests on the -bind address.
Boundary markers: None; the skill does not specify delimiters or warnings for instructions embedded in guest code or requests.
Capability inventory: The runtime supports writing to files (-logger, -kv, -secret-store) and making arbitrary network requests (-backend).
Sanitization: No sanitization or validation of the input WASM binary or request content is described.

fastlike