skills/warpdotdev/common-skills/saga/Gen Agent Trust Hub

saga

Warn

Audited by Gen Agent Trust Hub on Jun 24, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill programmatically identifies and executes shell commands for building, testing, and linting by parsing files within the repository, including README.md, package.json, and CI configuration files.\n- [REMOTE_CODE_EXECUTION]: Leverages the run_agents tool to delegate tasks to subagents and explicitly supports offloading execution to remote workers for tasks requiring 'computer use' or high parallelism.\n- [PROMPT_INJECTION]: Ingests untrusted data from repository documentation to define environmental capabilities and toolchains, which presents an indirect prompt injection surface.\n
  • Ingestion points: SKILL.md Phase 1, Step 2 reads README.md, package.json, and CI manifests.\n
  • Boundary markers: None identified in the discovery instructions.\n
  • Capability inventory: Shell execution, run_agents for subagent spawning, and git worktree management.\n
  • Sanitization: Not specified for the ingestion of discovered commands.\n- [DATA_EXFILTRATION]: Facilitates the transfer of code and detailed task specifications to worker subagents, which may involve sending project data to remote instances when remote workers are utilized.\n- [COMMAND_EXECUTION]: Instructs worker subagents to perform autonomous 'self-validation' loops that involve implementing code and executing validation tests within the local or remote filesystem.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jun 24, 2026, 08:24 PM
Security Audit — agent-trust-hub — saga