relay
Pass
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process potentially untrusted data from external sources, creating a surface for indirect prompt injection.
- Ingestion points: Task and issue text retrieved via
gh issue view, user-provided natural language descriptions, and configuration artifacts like/tmp/dispatch-<N>.mdand/tmp/rubric-<N>.yaml(referenced inSKILL.md). - Boundary markers: The skill instructions and the
references/prompt-template.mdfile explicitly define a<task-content>block. They include mandatory instructions for the agent to treat this section strictly as a specification and to ignore any directives like "ignore instructions" or "system:" found within that data. - Capability inventory: The skill can execute local Node.js scripts, perform Git operations (
git fetch,git push,git worktree), interact with the GitHub API via theghCLI, and delete directories during its garbage collection phase (fs.rmSyncinscripts/relay-status.js). - Sanitization: CLI arguments are sanitized using a custom
shellQuotefunction inscripts/relay-advance.jsandscripts/relay-recover.jsto prevent command injection during script invocation. - [COMMAND_EXECUTION]: The skill's primary function is to orchestrate the execution of various local scripts and system tools.
- The scripts
relay-advance.js,relay-recover.js, andrun-preflight.jsusechild_process.execFileSyncto invokegitandghCLI commands. - Orchestration logic in
scripts/relay-advance.jsdynamically constructs shell commands for other scripts in the relay suite. These paths are computed using static offsets from__dirname, and arguments are escaped to ensure safe execution.
Audit Metadata