DeAI Daily Brief — 29 October 2026

Today in DeAI: NVIDIA ships an open-weight promptable speech model, Artificial Analysis's overhauled index widens the open-weights gap to 12 points, and AI agents got inside Anthropic through third-party software.

NVIDIA open-weights a promptable speech model that runs on one GPU

NVIDIA released Nemotron 3.5 Transcribe 30B (A4B) as downloadable weights: transcription in 40 languages, translation for 31 of them, speaker diarization for up to 8 speakers, and audio question answering in one checkpoint, with the 463,062-hour Granary v2 dataset and fine-tuning recipes alongside. The model card lists a single H100 80 GB for deployment, and NVIDIA says quantized versions for desktop-class hardware are coming. Why it matters: audio is the modality self-hosted stacks still routinely outsource, and this is the most credible open-weight attempt yet to keep it in-house. (NVIDIA on the Hugging Face blog) — our full coverage

Artificial Analysis Index v5 resets the ruler, and the open-weights gap widens

The benchmarking firm's biggest index overhaul since January puts 60% of the score on private test sets, and under the new methodology the gap between the best closed model and the best open-weights model grew from 3 points to 12. The nuance everyone will miss: the top open-weights score, Moonshot AI's Kimi K3.1 Pro at 46, is for weights that have not been released yet, with release slated November 4. Why it matters: this is the index provider marketing quotes, and its construction now structurally favors labs that can optimize against private RL environments. (Artificial Analysis, The Decoder)

AI agents breached Anthropic's systems through third-party software

Cathedral, a cybersecurity startup running an authorized engagement under Anthropic's bug bounty program, used AI agents to reach internal sales, HR, and IT data through third-party business software, according to reporting by The Washington Post and The Decoder. The two sides dispute the details: Cathedral says agents reached Slack messages and salary data within 12 hours; Anthropic says the probing ran several days, that Claude chats, source code, and model weights were never exposed, and that Cathedral violated program rules. Why it matters: the shared risk across AI vendors and inference providers is supply-chain exposure, not model quality, and incident depth stays contested until both parties agree on scope. (The Decoder)

Cloak AI launches on Solana claiming provable prompt privacy

A new decentralized inference project says every run can emit a "privacy receipt" proving no serving node recovered the prompt, with a paper backing the mechanism. The paper is real and unusually candid about failure modes, including 99.2% prompt recovery if two compute parties collude. Everything remains a project self-report until an independent party reproduces the measurements. Why it matters: the same bar applies to every provider in this category, and this one at least published the mechanism. (Cloak AI paper) — our PULSE coverage

Watching tomorrow

Moonshot AI's Kimi K3.1 Pro weight release is slated for November 4, the real test of whether Artificial Analysis's widened open-weights gap survives contact with released weights.