Security
Don't trust us. Verify. Our security architecture is designed so you don't have to trust anyone.
Trust Chain
Hardware TEE (AMD/Intel/NVIDIA)
→ Open-source cage (Apache 2.0, you audit)
→ Audited engine (2 firms verified, signed report)
→ KBS attestation (binary hash verified before key release)
→ Your data: E2E encrypted, hardware-protected
Security Layers
Open-Source Cage
The security wrapper is fully open source (Apache 2.0). It controls ALL I/O between the engine and the outside world. Engine has zero network access, zero disk access. Enforced by seccomp-bpf + Linux namespaces.
Hardware TEE
All inference runs inside hardware-encrypted memory. AMD SEV-SNP, Intel TDX, or NVIDIA Confidential Computing. Not even the provider with root access can read your prompts.
E2E Encryption
Ephemeral AES-256-GCM session keys per request. Session key encrypted to TEE's attestation-bound public key. Platform routes encrypted blobs — cannot decrypt. Forward secrecy guaranteed.
Attestation + KBS
Key Broker Service verifies TEE attestation before releasing model decryption keys. Binary hash of cage+engine must match — any tampering = rejection. Re-attested every hour. Compromised provider = instantly offline.
Anti-Tampering
Steal binary + use own cage → hash mismatch → KBS rejects
Patch binary to skip KBS → hash changes → attestation fails
Extract key from memory → TEE encrypts memory → can't read
Run outside TEE → no attestation → KBS never sends key
Use with public unencrypted models → works but = plain llama.cpp, no advantage