The trust model
Understand the whole exchange.
An approval releases values or permits signing. Account permissions, the requesting computer, and the phone still matter after that decision.
- The phone
- Android, the app, and its authorized users control access. A compromised phone can expose values or authorize their use. SSH private keys are decrypted into app memory for signing.
- The computer
- Command details are reported by the client. They help you review a request, but a modified client can misrepresent them. Agentknock does not sandbox a command or enforce its behavior after approval.
- The relay
- Protected client–phone messages are encrypted end to end. The relay has no keys to read them, but it can observe routing information, sizes, and timing, and can delay or block exchanges.
- AI review
- Optional cloud review sends context to the review service and model provider. They can read that context and influence access decisions. The model can make a mistake.
Keep credentials out of review context.
Sensitive stored values and private keys are omitted from AI review. A credential typed into a command argument, reason, or instruction is still part of the submitted context.
Understand protection of past exchanges.
A stolen client pairing key alone cannot decrypt recorded past exchanges. A later compromise of the phone's protocol private key, combined with recorded pairing and exchanges, can expose them. Agentknock v1 has no forward secrecy against that device-key compromise.