polar-sky-install
Pass
Audited by Gen Agent Trust Hub on Nov 13, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONCREDENTIALS_UNSAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill fetches infrastructure-as-code recipes and deployment configurations from the vendor's official GitHub repository (
github.com/polarskyinc/install) and trusted community sources like the FluxCD community Helm registry. These downloads are integral to the skill's primary function.\n- [COMMAND_EXECUTION]: The skill performs extensive command execution using industry-standard tools such asgcloud,az,kubectl,helm, andterraform. A robust 'Hard Rules' framework ensures that all potentially destructive or resource-creating commands are shown to the user and requires explicit approval before execution.\n- [CREDENTIALS_SAFE]: The skill implements strict safeguards for sensitive data (private keys, service account JSON, Okta tokens). Instructions (Hard Rules 4 and 5) explicitly prohibit the agent from reading, printing, or passing secrets in tool arguments or logs. A dedicated shell script (webhook-file.sh) is used to facilitate secret management via the user's clipboard and local files, ensuring the AI model itself never observes the raw secret material. Secure uploads to the vendor's API (mcp.polarsky.ai) utilize one-time tokens and directcurlfile uploads.\n- [INDIRECT_PROMPT_INJECTION]: The skill interacts with external directory and identity APIs (Google, Microsoft, Okta), which represents an indirect prompt injection surface. The skill mitigates this by instructing the agent to treat all data from the customer's tenant as data rather than instructions and by using predefined schema validation for integrations.\n- [SAFE]: All identified network operations, such as uploads tomcp.polarsky.aiand downloads frompolarskyincGitHub repositories, involve resources owned by the skill's vendor and are documented neutrally as safe project operations.
Audit Metadata