Claude Duet, Granite Speech 5, Kumo Relational, attack surface

Four primary pages today: an agent app in preview, two open releases and a security write-up. I read each one and ran none of them.

Claude Duet

Anthropic describes Duet as a proactive agent. It monitors what needs your attention, works on tasks in the background and asks for approval before sensitive actions such as sending messages or making payments. It starts the day with a morning brief, takes larger jobs by message (meetings, dashboards, decks, spreadsheets) and shares memory, skills and apps with the main Claude app. The page cites more than 6,000 connectors and plugins and says it is powered by Opus 5.5.

It is a research preview behind a waitlist, open to Pro and Max subscribers in the US only, on iOS, Android and the web. Team and Enterprise are not included yet, and background checks count against your plan's usage limits. The page also admits that Duet may misread a thread, misjudge urgency or miss something that arrives between check-ins. That last point is the real test for any agent that acts without being asked.

Granite Speech 5.0

IBM released three speech models under Apache 2.0. The new granite-speech-5.0-2b transcribes English, French, German, Spanish, Portuguese and Japanese, translates to and from English, and reports 5.06% average English word error rate on the Open ASR Leaderboard short-form sets, second among top open-weight ASR models. It accepts keyword lists of up to 2,000 terms.

The 470M turboctc model is now multilingual and reports 19,870 RTFx, up from about 12,600 at its August launch at the same accuracy. The 4.1-2b-plus update targets meetings and calls: speaker-attributed WER moves from 30.50 to 28.48 and biased-keyword WER from 11.79 to 6.48, as a drop-in replacement.

NVIDIA Kumo Relational

Kumo Relational predicts outcomes like churn, demand or fraud straight from a relational database, with no task-specific training or table flattening, using up to 10,000 historical examples as context. NVIDIA reports results on 41 tasks across 15 databases that were unseen in pre-training, including about 4 AUROC points over gradient-boosted trees on RelBench V1 classification. Code is Apache 2.0 and weights are on Hugging Face under the OpenMDW license. The authors note that supervised models may still win when the training set is very large.

Reducing the attack surface of code

Anthropic's Frontier Red Team used models to map which code is reachable and cut what a deployment does not need. Over 25 projects and more than 5 million lines, it reports 493 issues, including 162 security-sensitive option combinations accepted with no warning. In OpenSSH 10.4, about 4,500 lines of experimental ML-DSA-44 code were reachable by unauthenticated clients in the default configuration, and the fix in 10.6 removed roughly a fifth of that reachable code.

What I'd watch next

Whether Duet widens beyond the US waitlist, and how its approval prompts hold up in real use. Whether Kumo's gains survive on my own messy tables. And whether the attack-surface skill, headed into OSS-Scanner, finds the same kind of gaps in other projects.