Security: w8mej/Mimir
Security
SECURITY / THREAT MODEL (MVP)
Demonstrate confidential inference workflows for autoregressive models.
Preserve input and model privacy under honest-but-curious assumptions while enabling token-by-token generation.
Provide hooks for auditability and future attestation.
Honest-but-curious 3PC : at most 1 colluding party among 3.
Nodes run in controlled environments; containers provide namespacing but no strong isolation.
RNGs are cryptographically strong when upgraded (current MVP uses NumPy RNGs).
Inputs : prompts/token IDs
Parameters : model weights
Intermediates : attention scores, activations, KV-cache
Outputs : generated tokens
Secret sharing (replicated) for tensors and weights.
Some intermediate opens for practicality (attention scores, V-context, softmax normalizer).
KV cache stored publicly for demo. (Upgrade path: keep K/V secret-shared.)
Transport security is not implemented in code (gRPC insecure channels for local dev).
Prometheus metrics are non-sensitive counters/histograms only.
Information leakage via opened intermediates.
Side channels through timing/size patterns if deployed cross-tenant.
RNG predictability for sampling variants.
Replace opens with masked reveals + Beaver-based secure matmuls for QKᵀ and AV.
MACed shares (SPDZ) for malicious security; sacrifice checks for triples.
mTLS between nodes + optional WireGuard overlay; pin container digests.
Remote attestation (AMD SEV-SNP / Intel TDX) with quote verification in Coordinator.
Differential Privacy at logits layer; configurable ε, clipping.
Secret-shared KV cache with per-step garbage collection and transcript hashing.
There aren’t any published security advisories
You can’t perform that action at this time.