The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill requires the agent to verify bug reports by reproducing them from reporter-provided steps and to verify pull requests by running tests or commands against the external diff. This constitutes execution of untrusted code from unknown sources.
Evidence: "For a bug, reproduce it from the reporter's steps."
Evidence: "For a PR, confirm the diff does what it claims: check it out, run the relevant tests or commands."
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external issues and pull requests, creating a surface for malicious instructions to influence agent actions.
Ingestion points: Issue body, PR body, comments, and code diffs in SKILL.md.
Boundary markers: No delimiters or explicit warnings are used to isolate untrusted content from instructions.
Capability inventory: Shell command execution (tests/commands), file system writes (.out-of-scope/*.md), and GitHub API operations.
Sanitization: There is no evidence of content filtering or validation before the agent acts on external inputs.
[DATA_EXFILTRATION]: The ability to run arbitrary commands during verification combined with access to the codebase and environment creates a risk of data theft if an attacker provides malicious reproduction steps.
[COMMAND_EXECUTION]: The skill encourages the agent to explore the codebase and run verification tests using shell commands without implementing a restricted or sandboxed environment.

triage