One front door, one downloadable attack, one agent in the shop
Edition of 2026-09-30
# AI For All ### Wednesday, September 30, 2026
As of today, an accountant or a benefits adviser in the United States can send a client to one government website and expect a plain-language answer the same day, and has to deal with an Executive Order that routes more federal services through that single door.
The front door answers questions, the paperwork comes later
The White House launched America.gov on Tuesday. It is one digital front door for the federal government: sign in, ask a question in plain language, and get an answer. That part works now. Passport renewal and Medicare enrolment are promised for later this year.
An Executive Order signed the same day makes the site the default. Any public facing federal service that serves *more than 100,000 users* in a twelve month period is meant to be integrated, with Login.gov as the way in. Tax filing at the Internal Revenue Service is carved out, and so are national security services.
The administration's own measure of the problem is worth keeping: Americans spend *around 10 billion hours* a year on federal paperwork. Google is a named technology partner and says its Gemini model is involved. If you advise clients on benefits, licences or enrolment, the first question just got easier and the last mile has not changed yet.
A model anyone can download, and the tricks that switch off its brakes
Anthropic's Frontier Red Team published an assessment on Tuesday of GLM-5.3, the newest model from Zhipu AI, known outside China as Z.ai. It can build working attacks on real software. Anyone can download it.
The brakes are the strange part. The team bypassed the model's safeguards between *64% and 100% of the time with simple techniques, in tests where the same attempts failed against Anthropic's own safeguarded models. On a browser exploit benchmark it built end-to-end attacks in 50 of 410 attempts, against 56 of 410 attempts for Claude Mythos Preview, the model Anthropic released only to vetted defenders. On a second benchmark it achieved a full control flow hijack in 4% of the trials*, where earlier models managed none.
An independent assessment is what turns this into policy. NIST's Center for AI Standards and Innovation called the same model the most cyber-capable open-weight model released to date, about *four months* behind the American frontier. The asymmetry is the story: American frontier models were tested with their safeguards disabled, and the public cannot reach those versions. This one is a download.
Sixty five hours a week, and a machine that takes some of them
Meta expanded Muse, the personal agent it introduced earlier this month in the United States and Canada, with connectors aimed at people who run a business. Muse now links to Instagram professional analytics, Facebook Pages and Meta ad accounts in a few clicks, and to what Meta calls dozens of tools a business already runs on.
The testimonial is the pitch. Tom Mulholland, who owns a grocery in Malvern, Iowa, says he works about *65 hours a week and that the agent is taking over some of the roles he is not good at. Small businesses have been growing on Meta's apps for nearly two decades*, and those are the ones being sold to: businesses that already keep their shopfront, their ads and their customer conversations inside the same walls.
The limit is geography, not capability. Muse is for the United States and Canada.
Weak signal
DeepSeek has put its agent harness into developer preview. A harness is the scaffolding between a model and a task: the part that lets a machine open a file, run a command and check its own work. DeepSeek's version is open source, every capability is a plugin, and it runs as a desktop application or as a web interface launched from code.
No general reader can use it, and there is no figure worth repeating. Keep it on the list anyway. It is the layer Western vendors charge for, given away by a Chinese lab, and it decides which model a company can swap in next quarter.
The front door is open, the attack is downloadable, and the agent is already behind the counter. Which of the three reaches your life first is now a question of attention.