Assume everyone is watching. Design so it doesn't matter.
DarkGPT assumes the provider is curious, the network is observed, and the chain is public. Here is exactly what each of them can see.
Who sees what
Every peer connection is authenticated and encrypted end to end with ephemeral session keys, so a captured session cannot be replayed or decrypted later.
Wallet and decryption keys are generated inside the enclave with hardware attestation. They never touch the host OS, and high value transfers require biometric approval.
Inference runs in an isolated VM against a hash verified Ollama build. Plaintext exists only inside that boundary, and only for the life of the request.
Each escrow carries a unique nonce and each proof can be spent once, which closes double payment on both sides. Unproven work refunds automatically at ten minutes.
What we do not claim
Honest limits, because a security page that only lists strengths is marketing, not a threat model.
Report what you find
Found a weakness? Send disclosure reports to the security address and they are triaged within 48 hours. Every build is signed, notarised by Apple and integrity-checked on launch.