The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection as it ingests untrusted data from docstrings and comments in the authorized source tree (defined in SKILL.md). Boundary markers are present in the form of explicit instructions to treat such text as untrusted data. The capability inventory includes the ability to execute repository-native checks and modify source files. Sanitization is achieved through instructions to redact secrets and avoid reproducing unnecessary private text.
[COMMAND_EXECUTION]: The skill leverages existing repository-native tools (e.g., go test, cargo test --doc, dotnet build) as described in references/protected-comments-and-native-checks.md to verify changes to protected content like doctests. Security is maintained by requiring explicit user authorization for these commands and forbidding the installation of new dependencies or unauthorized toolchains.

comment-health