DeAI Daily Brief — 4 November 2026

Today in DeAI: Moonshot's Kimi K3.1 Pro weights deadline arrives without a repo, Anthropic publishes a zero-knowledge model-provenance prototype, AWS agent security gets a rough disclosure, and Sarvam ships a small Apache 2.0 decision model.

Moonshot's Kimi K3.1 Pro weights deadline is here, and the repo is not

The launch post's sentence, "Full model weights will be released by Nov 4, 2026," is still on Moonshot's page on deadline day. As of about 04:20 UTC, the moonshotai Hugging Face org still lists Kimi-K3 from June as its newest model, and the only third-party K3.1 repo is gated and empty. Why it matters: self-hosting and provider plans for a 2.8-trillion-parameter model hang on this drop, and the unannounced license is the real open question. (Moonshot AI) — our full coverage

Anthropic posts FlashProve, a zero-knowledge proof that a reply came from committed weights

The FlashProve repository, created October 30 under Apache 2.0, lets a provider commit to a weight set and produce a zkSNARK proof per reply, verifiable on a CPU in seconds. The repo reports running Llama 4 Maverick at a 1,024-token context on one H100 at about 110 tokens per second of prefill, and it self-describes as an unmaintained research prototype, not a product. Why it matters: cryptographic model provenance is the missing trust layer for hosted inference, where silent checkpoint swaps are currently taken on faith. (GitHub) — context in our provenance coverage

Zenity Labs' AgentCorruption: one prompt injection reached AWS account credentials

Zenity Labs' SecTor 2026 research showed a single prompt injection to a public-facing Amazon Bedrock AgentCore agent could reach the instance-metadata endpoint and harvest execution-role credentials, exposing secrets across every agent in the same account and region. AWS required IMDSv2 for new deployments in February and removed the broad permissions by September 29, and disputes calling documented IMDS behavior a vulnerability. Why it matters: agent platforms inherit cloud IAM blast radius, so per-agent role scoping is a checklist item, not a nice-to-have. (Cyber Security News) — related: the Cloudflare KV-cache cross-tenant leak

Sarvam AI ships Prism, an Apache 2.0 decision model for routing and classification

The sarvamai/prism repository landed November 2 with Apache 2.0 licensing and a text-classification pipeline tag covering English plus nine Indian languages, built on Sarvam's 105B foundation line. Claimed benchmark and latency figures circulate from the launch post and are not on the model card; treat them as vendor claims. Why it matters: a small open router model can replace frontier-API calls in the high-volume middle layer of a stack, and Apache 2.0 makes it directly self-hostable. (Hugging Face)

The privacy terms divergence widened: OpenAI ZDR, Anthropic's 30 days, Cohere's vault

OpenAI's Zero Data Retention expansion to frontier models under the "Private Intelligence" umbrella landed alongside Cohere's Model Vault encryption push, with coverage framing both against Anthropic's 30-day retention default. Every "zero retention" and "cannot see" characterization in this cluster is a vendor claim pending independent attestation. Why it matters: retention windows now differ enough across labs to be a real selection criterion when choosing where to run models. (VentureBeat) — our ZDR explainer and Private Intelligence preview coverage

Watching tomorrow

The Kimi K3.1 Pro repo: whether it lands late today, slips quietly, or appears with a restrictive license decides the week's biggest open-weights story.