Kruvo
Security

Security policy

Kruvo runs on your machine with your model keys, your connector tokens and access to your files and terminal. Any report that lets an agent, a web page, a message or a tool escape its approval gates is a security issue to us.

Reporting a vulnerability

Email [email protected]. Include the Kruvo version (Settings → About), the operating system, the connector or provider involved, and steps that reproduce the issue. Do not paste bot tokens, API keys or private conversation content; redact them or describe where they appear.

You get an acknowledgement within 3 business days and a fix or mitigation timeline within 10. We credit reporters in the release notes unless you prefer otherwise. Machine-readable contact: /.well-known/security.txt.

In scope

  • Approval bypasses: an action that should ask a human and does not.
  • Hard-floor bypasses: writes to protected files, persistent-authority tools, MCP trust.
  • Prompt injection that reaches a tool call with attacker-chosen arguments.
  • Secrets leaking into logs, transcripts, artifacts or messages.
  • Path traversal, SSRF through fetched URLs, unsafe exposure of the local API.
  • Installer, updater and signing weaknesses; the kruvo.ai services.

Out of scope

Issues in the models themselves, in third-party services you connect, or in dependencies already disclosed upstream without a Kruvo-specific impact.

How Kruvo is built to be safe

  • Hard floors that no mode can lower.
  • A reviewer model that can only turn "ask" into "allow", never "blocked" into "allow", and fails closed.
  • A circuit breaker that pauses auto-approval after repeated denials.
  • An audit log with secrets redacted, persisted with the conversation.
  • Signed auto-updates: the app installs only manifests signed with the Kruvo updater key.