The loudest security story of the day was T3MP3ST, released by jailbreak researcher Pliny: a multi-agent offensive framework that plugs into your existing Claude Code, Codex, or Hermes login and drives real recon — nmap, DNS, HTTP probes — through the agent. What made it spread was that every finding is provenance-gated (you can re-derive each number with a verify-claims command) and kill-chain phases past recon are labeled honestly instead of faked. It reframes a coding-agent subscription as a full red-team console.

International Cyber Digest laid out the framing, and the rest of the feed was unusually concrete: Strix packaged pentest agents you point at your own app, a 1.5B local security model did Blue/Red-team reasoning on CPU, a warning landed that every unknown "AI browser" is a prompt-injection liability, a "Bad Epoll" Linux/Android root 0-day made the rounds, and Stanford's Real World AI Security Conference showcased fresh attack research.

A recurring thread was that agent skills stopped being private config and became public artifacts. David Ondrej dumped every global .agents skill he had — "hundreds of hours of trial and error" — for anyone to grab, and it was quickly echoed around the timeline.

Even Google's Addy Osmani open-sourced his whole Claude setup, others pointed to standout skill repos, and people traded concrete workflow use cases like candidate sourcing. The payoff is real — skills are becoming the scaffolding for "1-person companies" of coordinated local agents — which is the productivity flip-side of the same skills-as-supply-chain surface the security crowd keeps flagging.

Days after launch, the Fable 5 conversation shifted from "new model" to "what workflow do I wire it into." The viral claims were about compressing agency-scale work into a single session — an animated website, a shippable app, a game done in one run for a few dollars of credits. Take the dollar figures with salt, but the pattern is real.

One claim put an agency-scale animated site at ~$12 in credits, and another shipped a full iPhone app in a single run. The more useful tweets were about orchestration: people converging on a Hermes-manages / Fable-plans / Codex-builds split, running Fable 5 against Opus 4.8 on identical briefs, noting Fable's product-level judgment, and cataloguing which coding front-end still lacks Fable support.

Local AI chatter has fully moved off "can it run locally?" and onto the buy-vs-subscribe math. The framing tweet was blunt — $300/month in subscriptions, or one ~$1,500 box under your desk that pays for itself in five months — pinned to Microsoft's mini-PC AI dev-box reveal.

The evidence came from the homelab crowd: someone ran the full unpruned 744B GLM-5.2 across four DGX Sparks, another stacked Sparks to keep two models resident at once, one maxed out a fully local agent-plus-robot stack, and a builder assembled an AI server entirely from spare parts. It ties back to the Qwen lead's talk on where open agent models are heading.

The practical AI-video story was Seedance 2.5 landing inside consumer editing tools rather than staying an API novelty. CapCut announced Dreamina Seedance 2.5 with seamless generate-and-edit, 30-second scenes, and up to 50 multimodal references — the reference-count leap is what lets creators hold a character or product consistent across a whole spot.

Lumina echoed the same 30-second, 50-reference jump, while supporting posts already showed the output — 4K product commercials built end-to-end from Seedance and a preview of the 50-asset customization on CapCut.

A standalone security note worth flagging: a verified X account ran a sponsored post promoting what looked like DynamicLake — the popular Mac notch/utility app — but the download was Mac malware. A clean reminder that verified badges and paid placement are not trust signals, and that installing Mac utilities from ads instead of the developer's own site is a live way to get burned.