Anthropic Flags GLM-5.3's Exploit Skills in a Cyber Report

NVIDIA published OpenShell 0.1.x, a runtime that gives fleets of autonomous agents file, package and API access under a declared policy. Each agent runs in an isolated sandbox; the kernel confines its file access and system calls, and every network connection passes a policy check before it leaves. Agents never see real credentials, which OpenShell adds only to requests bound for approved endpoints. A companion walkthrough runs OpenCode against a free OpenRouter model and shows how to approve new access as the agent needs it.

@poteto published part two of his pstack guide, the workflow behind 2,462 pull requests in production at the end of August. The /why skill queries Git history, Linear tickets, Notion docs, Slack threads and Datadog monitors in parallel. /architect drafts competing designs across model families and judges them against a rubric. He plans by prototyping instead of writing plan documents, keeps 23 playbooks in /poteto-mode as of 0.15.0, and shares his most used prompt: restate my goals in your own words.

@shadcn argues that apps like ChatGPT, Claude Code and T3 Code should keep live threads in horizontal tabs at the top of the window. Vertical tabs are a tree of history you rarely reopen; the working set wants icons, colors and status at a glance. His wishlist post to the ChatGPT team asks for “horizontal tabs. trust me.” @frogggias counters that multi-agent work needs project, task, status and machine, more fields than a tab carries.

@mmastrac calls kindlingai/glm-5.3-flash-gx10 the fastest way to run GLM Flash 5.3 on 2, 3, 4 or 6 Sparks, with switch or switchless TP=4RING topologies. At TP=4 the repo measures 114.6 tok/s code decode, 4,981 tok/s cold 32k prefill and needle recall of 12/12 up to 507k tokens, with 524k context per request. DFlash2 speculative decoding and NVFP4 MoE kernels supply the speed; FP8 copies of the dense layers cost about 1% NLL on prose. On their switch, the switchless relay hop cost about 1% of decode.

@KennethCassel reports his supplier’s warning that data centers are buying up metal stock. @rossiadam says a New Jersey supplier quoted 12 weeks for structural steel with all available output going to data centers. @GonsiorJacob says an aluminum order came in 25% above its quote from two weeks earlier. @SMBAwningMan says his powder coater is booked with data center work, and @DKFlii adds that wire and cable are hard to get.