The confidential inference weekly
Attested
What changed in confidential inference — and what the evidence actually shows.
Subscribe on SubstackIssue #4 ·
Attested #4: Anthropic can prove which model answered. Not who read your prompt
Two provider breaches and what an enclave changes, FlashProve proves the model but not your privacy, Kimi K3.1 Pro on six services with six meanings of private, and a CVE for diversion attacks
Issue #3 ·
Attested #3: OpenAI's "confidential computing" starts after it reads your prompt
OpenAI's real-time safeguards read your prompt before Private Inference starts. Plus abliterated models in enclaves for security teams, Chutes cuts two TEE prices and NEAR moves DeepSeek 1017 into its own TEE
Issue #2 ·
Attested #2: Phala publishes its dstack post-mortem, DeepSeek 1017 spreads (not always into an enclave)
The TEE held and the plumbing around it didn't. Plus why "live on NEAR AI Cloud" doesn't mean attested, Cloudflare's cross-customer KV cache leak, and Chutes picks up post-quantum TLS
Issue #1 ·
Attested #1: Brave puts Leo behind attestation, Tinfoil publishes a measured build of DeepSeek's newest weights
This week in confidential inference: attested Leo, hash-pinned DeepSeek, a 38% Kimi K3 cut, and 0 of 97 monitored checks verified end to end