openshell-cli

Pass

Audited by Gen Agent Trust Hub on Nov 20, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONPERSISTENCE
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to monitor sandbox logs using 'openshell logs --tail' to identify 'DENIED' actions and refine security policies. This presents a surface where malicious output from the sandbox could influence agent behavior.\n
  • Ingestion points: Log output from the sandbox main process (SKILL.md).\n
  • Boundary markers: None provided; instructions suggest direct parsing of log lines.\n
  • Capability inventory: The agent can execute commands, modify security policies, and manage credential providers via the CLI.\n
  • Sanitization: No sanitization is mentioned for processing external log data.\n- [EXTERNAL_DOWNLOADS]: The skill documents the ability to import or lint sandbox profiles from remote HTTP/HTTPS URLs.\n
  • Evidence: 'openshell profile import --url https://example.com/profiles/my-profile.yaml'.\n
  • Mitigation: The skill explicitly warns to 'Review its endpoint and binary grants before importing it.' and limits downloads to 1 MiB.\n- [COMMAND_EXECUTION]: The agent is guided on how to execute arbitrary commands within an isolated sandbox environment.\n
  • Evidence: Usage of 'openshell sandbox exec' and 'openshell sandbox create -- '.\n- [PERSISTENCE]: The skill includes a workflow for modifying the local user's SSH configuration file to facilitate persistent connectivity to managed sandboxes.\n
  • Evidence: 'openshell sandbox ssh-config my-sandbox >> ~/.ssh/config'.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 20, 2026, 07:24 PM