golive
Pass
Audited by Gen Agent Trust Hub on Nov 4, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill's installer (
scripts/install-lib.mjs) fetches release metadata and file bundles from the author's GitHub repository. These downloads are restricted to version-controlled tags and are verified for integrity. - [REMOTE_CODE_EXECUTION]: The installer downloads and executes skill scripts. This process is secured by SHA256 integrity verification for all files and a 'smoke check' sandbox that disables network access and subprocess spawning during the initial verification of new versions. It also explicitly warns the agent against executing remote scripts for third-party CLI installations, directing those tasks to the human.
- [COMMAND_EXECUTION]: The skill invokes local CLI tools and Node.js scripts to manage infrastructure and verify deployments. It includes detailed instructions for the agent to avoid unsafe command patterns and maintain least-privilege operations.
- [DYNAMIC_EXECUTION]: The skill uses
node:child_processto run its internal CLI for deployment tasks and verification. It also dynamically generates a temporary 'guard' script to sandbox the environment during updates, preventing network and process spawning by patching core Node.js modules likenode:httpandnode:child_process. - [INDIRECT_PROMPT_INJECTION]: The skill implements a defense against indirect prompt injection.
- Ingestion points: Processes untrusted repository data (READMEs, config files) and provider API responses during the detection and planning phases.
- Boundary markers: Explicitly instructs the agent (Rule 7) to treat external content as data rather than instructions, requiring manual reporting if command-like content is detected in data.
- Capability inventory: Performs file system writes, network requests to cloud providers, and infrastructure management operations.
- Sanitization: Relies on structured data exchange and requires explicit human approval for all high-risk writes and destructive operations.
Audit Metadata