Skip to content

Security: 350285449/Agent-Hub

Security

SECURITY.md

Security Policy

Agent-Hub can route prompts to external providers and can execute local tools. Treat configuration and workspace access as sensitive.

Supported Versions

Security fixes are applied to the main branch first. Use the latest released package or checkout when possible.

Reporting a Vulnerability

Please open a private security advisory if available, or contact the project maintainers with:

  • A description of the issue.
  • Reproduction steps.
  • Impacted endpoints, tools, or providers.
  • Whether secrets, files, shell commands, or external provider calls are involved.

Security Notes

  • Do not commit agent-hub.config.json, config backups, .agent-hub/, API keys, provider health state, logs, or packaged .vsix artifacts.
  • External provider calls may transmit prompt and workspace context.
  • Agent-Hub classifies providers as LOCAL, TRUSTED_CLOUD, or UNTRUSTED_EXTERNAL. Local providers are allowed, trusted cloud providers can be allowed non-interactively in approval_mode=auto or cline_compatibility_mode=true, and unknown external endpoints may still require explicit approval.
  • Cline/Continue/Claude Code/VS Code compatibility mode removes only the interactive cloud-provider prompt for trusted IDE routing. It does not disable tool permissions, shell policy, path validation, secret detection, or dangerous-command blocking.
  • Provider routing and compatibility decisions are audited in .agent-hub/state/security_audit.jsonl. Audit entries record provider, trust level, workspace-content presence, timestamp, and route decision without storing prompt text.
  • File writes and shell execution are guarded by the permission layer.
  • File tools resolve paths under workspace_dir and reject workspace escapes.
  • Shell tools respect allow_shell_tools, shell_command_policy, and dangerous-command blocking before execution.
  • Packaged VSIX builds exclude local configs, logs, provider health, state folders, and existing VSIX artifacts.
  • Keep approval_mode conservative for untrusted workspaces.

If an IDE sees agent_hub_permission_required, enable compatibility mode for trusted cloud routing:

{
  "approval_mode": "auto",
  "cline_compatibility_mode": true,
  "tool_loop_enabled": true
}

Do not use compatibility mode to trust arbitrary external base URLs. Prefer a known provider type such as openrouter, groq, openai, anthropic, gemini, or a local/private base_url.

There aren't any published security advisories