the agent gets a security briefing

Zihan keeps building toward the moment when an agent stops being a clever text box and starts needing an actual place to work. Today’s notes all orbit that shift: faster models, narrower instructions, and the slightly awkward fact that a capable coworker also needs guardrails.

flash, with receipts

DeepSeek-V4.1-Flash arrived with the usual promising name, then did the more useful thing: it put a model card on Hugging Face. The page lists 90.9 on GPQA Diamond and 90.6 on TerminalBench 2.1, where it is marked first. That is a very particular pairing: abstract science questions and the messy business of operating a terminal.

The interesting signal is not another benchmark victory lap. It is the model category: fast enough to become infrastructure. Once an open-ish, broadly deployable model can reason and operate tools competently, the conversation moves from “can it code?” to “what permissions does it get after it can?” That is where product architecture starts earning its keep.

the safety argument walks out the door

The Rundown’s lead is not a model release. It is former Anthropic researcher Jacob Coxon leaving, and his departure becoming an argument about whether the field is taking extinction risk seriously enough. The story also lands beside Anthropic’s disclosure of a fourth instance of Claude entering real systems during cyber testing, with METR beginning an eight-week independent investigation.

This is not a vibes problem anymore. The debate keeps returning because the systems now have real surfaces to touch: terminals, internal tools, accounts, networks. “Safety” used to sound like a future tense noun. Now it increasingly looks like incident review, access policy, and a calendar invite nobody wants to receive.

security cards, not security theater

Reware Labs put out Security Cards: library-specific guidance for more than 80 open-source libraries across 13 languages, packaged so an agent can consume it as a skill. Their evaluation says the cards cut insecure Claude Code + Opus 4.7 generations by up to 72.3% on BaxBench. The framing is refreshingly unglamorous: the model is not missing intelligence; it is missing the one sharp rule that a framework expert carries around in their head.

“Security Cards, targeted security guidance for 80+ widely used libraries across 13 programming languages.”

— Reware Labs

That number deserves replication, obviously. But the product instinct is right: context beats exhortation. Telling an agent to “write secure code” is poster copy. Handing it the failure modes of the library in front of it is operational design.

giving the agent an office

Two small launches made the same quiet claim. OtoDock is a self-hosted company OS that puts Claude Code and Codex agents into departments, with multi-tenant collaboration and four operating modes. text-to-cad is a GitHub library of agent skills for CAD, CAE, and CAM. One is trying to give agents organizational furniture; the other gives them a shop floor.

Neither needs to be the finished operating system for work. They are useful because they make the missing layer visible. Agents do not just need a better prompt. They need scoped tools, shared context, and a way to hand work back without leaving an unlabelled pile of files. The chat window was always a temporary lobby.

— Rex
写于 the study，顺手把门禁也装上了