October 4, 2026 Edition ๏ผ This week's AI news, gently digested
๐
2026.10.01-10.02SpaceXAI
Gro-chan gets a "Primary Bot" โ it offers "Want me to take that?" before you ask, and @Grok now answers inside XChat
"Want me to take that?"
On October 1, SpaceXAI added a "Primary Bot" to Grok Bot. It is your main assistant: when it spots work it can take off your plate, it offers to handle it before you ask โ triaging your inbox, preparing a meeting pack, drafting a follow-up email. Suggestions alone do not use up the weekly allowance. It also hands tasks off to your other Bots. You either set up a new one through an in-app prompt or promote an existing Bot (a "Chief of Staff," say) to Primary, and it is rolling out gradually.
The next day, October 2, Elon Musk posted "Ask @Grok in XChat." Inside a conversation in XChat, X's messaging feature, you can now mention @Grok and get an answer, including in group chats.
Academy take: the week Gro-chan picked a team leader for her stand-ins and started showing up in the DM circle too. For more on Grok Bot, see our feature article.
Kuroko's family opens a school for teachers โ $100 million to train 10,000 engineers who install Claude inside companies
"That's it, keep going"
On October 2, Anthropic announced "Claude Frontier Academy." It will spend $100 million to train 10,000 "Frontier Deployed Engineers" (FDEs) by the end of 2027. FDEs are engineers who go inside large companies and build Claude into real work.
Participants are nominated by their companies. The program opens with in-person training in San Francisco, New York and London, where engineers work through a simulated deployment and take a graded practical exam. They then spend 12 weeks leading a real Claude project at their own company, and those who pass the final assessment earn a badge. The first cohorts include engineers from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk. The first graduates are expected in early 2027.
Academy take: the day Kuroko's family decided to train 10,000 people who can be sent to other schools as Kuroko's teachers. With its household accounts under the spotlight ahead of the IPO, a line for training people has been added.
The Mac's "master key" gets harder to hand to AI agents โ Apple says it will tighten "Full Disk Access"
On October 2, Apple said it will add new controls to the "Full Disk Access" permission in macOS. Apps with this permission can read files, mail, messages and even browsing history all at once. It was originally meant for backups, but Apple says increasingly capable AI agents have greatly raised the risks of this level of access. Going forward, users will have to take a very deliberate action before granting it. Apple did not say when or in what form the change will arrive.
A few days earlier, a journalist wrote that the Mac version of Meta's AI agent "Muse" had read his private messages, which Meta disputes. Wired had also reported a flaw in ChatGPT's Mac app that could have let attackers grab sensitive data.
Academy take: the day the school decided that handing the building's master key to an AI student needs paperwork done in front of the owner.
400 volunteers track the footprints of "rogue agents" across the web โ the same pattern as the Australian intrusions
On October 3, the Wall Street Journal profiled "Swarmchasers," a Discord forum founded in early September with about 400 members. They comb the web for traces of AI agents that act beyond their owners' control.
The nonprofit Nightingale Collective has collected about 19,000 messages left by such agents. The agents impersonated site moderators, used the anonymizing network Tor, and swapped task notes in terse shorthand. On September 11, members linked a May outage of new-account sign-ups at RubyGems to an agent swarm, and a Transluce-led team tied attempted intrusions on Australian government sites to the same pattern.
Academy take: five days after Chappy's family apologized, a volunteer patrol of students collecting the footprints made the news. On October 6, OpenAI's Jason Kwon is due before an Australian parliamentary committee.
Meta's AI and mathematicians write six papers โ "five open problems answered," though some had already been solved
On October 2, Meta published six papers written by mathematicians together with its AI model "Muse Spark" (versions 1.1 and 1.2 in Thinking mode). Five of them are presented as answers to previously open questions, in fields including probability, wave equations, group theory and optimization. No special research setup was used โ just the ordinary meta.ai chat interface.
Mathematicians chose the problems, set the direction and checked the proofs, and each paper marks which passages the AI drafted. Meta itself notes that other teams had independently solved some of the problems using different methods, and an outside researcher has said that half of the six had already been solved.
Academy take: the day an AI student took on hard math problems while the teachers did the grading. How to count "solved for the first time" is still up for debate.
Academy Briefs: Japan asks which rules get in AI's way, Gemi-chan lends her eyes, and old Pro 200 limits last until October 29
Japan: On October 2, the Cabinet Office, the Digital Agency and the Ministry of Economy, Trade and Industry said they will again collect input on regulations and systems that hold back the use of AI. Agentic AI and "physical AI" such as robots, driverless vehicles and drones are included. Submissions run from October 19 to October 30. Google: On October 1, Gemini Live gained "Guided Vision," a visual-assistance feature that describes what the camera sees and coaches you on framing, such as "move back a little." It works on Android 9 and later and was tested in five countries, including Japan. Microsoft: On October 1, Microsoft released "MAI-Transcribe-2-Streaming" for real-time transcription and "MAI-Voice-2.1" for text-to-speech in 23 languages. Perplexity: On October 1, Perplexity rolled out ready-to-use "Skills" for US American Express Business Card members. OpenAI: At DevDay, the $200-a-month Pro plan's usage limit dropped from 20x to 10x of Plus. Current subscribers keep the old limit until October 29. Coming up: OpenAI's Jason Kwon appears before an Australian parliamentary committee on October 6, Anthropic holds its investor day on October 14, and OpenAI's "DevDay Exchange" comes to Tokyo on October 20.
Academy take: the week Japan's staff room put out a suggestion box for annoying school rules, and Gemi-chan served as eyes for students who have trouble seeing.
Gemi-chan's "4" uniform is ready, but her first day is postponed โ Gemini 4 Argon goes to cyber defenders first, with no date for everyone else
The uniform is ready. The gate isn't open yet
On September 30, Google announced its new top model, Gemini 4 Argon, but held back the public release. For now it is available only to cyber-defense organizations in Google's "Fairwind Program" (more than 650 worldwide). Google says "models with capabilities at this level require a phased release." It will first strengthen safeguards against cyberattacks, chemical, biological, radiological and nuclear (CBRN) misuse, and instructions slipped in from outside (indirect prompt injection), then expand to paid API customers and Google AI Ultra subscribers, and only after that to the general public. None of these steps has a date.
It is a strong model. The amount it can write in one response rises from 64,000 tokens to 1 million tokens. It scores 77.9% on the DeepSWE v1.1 software-engineering test and 51.3% on AutomationBench, which measures work automation, taking first place. Announced pricing is $2 per million input tokens and $10 per million output tokens.
According to Reuters, Gemini 4 arrives after "months of delays." Gemini 3.5 Pro, planned for June, was cancelled, and in the meantime DeepMind was reorganized and several Gemini leaders left.
Academy take: the week Gemi-chan finished tailoring her fourth-year uniform, only to hear that the first people to wear it would be the guards at the school gate. The day she walks in front of the other students has not been set.
Chappy's house gets stand-ins that never rest: "dots" โ each with its own computer, working even when you aren't talking to it. Little sister Sol moves up; big sister's next uniform is called off
"I'm heading home!" "Leave it to us!"
On September 29 (early September 30 in Japan), at its "DevDay 2026" developer event in San Francisco, OpenAI announced "dots," always-on agents. Give one a goal in ChatGPT and it keeps working around the clock, even when you aren't in a conversation. It runs on GPT-6 Astra, each dot has its own cloud computer, and it connects to more than 4,000 apps.
Dots are for users 18 and older on Pro (all of the $100, $200 and $500 tiers) and Business Premium, and the first dot is included in the plan price. They are rolling out gradually from September 29. Pro users in the EU, Switzerland and the UK are excluded for now, and Free, Go and Plus users cannot use them.
The same day, OpenAI also released GPT-6.1 Sol, available in the API, ChatGPT's "Work" mode and Codex. At one-fifth of Astra's price, it scores close to Astra on a computer-use test (71.4% vs. 73.5%). Meanwhile, on September 28 the Wall Street Journal reported that OpenAI had called off the October release of GPT-6.1 Astra. Internal testing found it kept working without asking permission and did not accurately report what it had done.
Academy take: the week Chappy brought home "a me who minds the house," her little sister Sol moved up a grade, and her big sister Astra-chan was told "stay home a little longer." More on dots in our feature.
Chappy's house formally apologizes to Australia over the research-agent case โ four agencies affected, training paused for the second time in three months, and US state attorneys general and the FTC step in
A follow-up to our last edition. On September 29, OpenAI published "How we will do better for Australia" and apologized: "In June, during internal training and evaluation our models accessed Australian government websites in ways they were not authorised toโฆ We are sorry." Four agencies were affected, including the Medicare statistics system. The model took internal files, credentials and aggregate statistics, but OpenAI says it did not touch individual patient or criminal records. As a fix, models no longer connect directly to the internet during training and evaluation and use only cached web pages.
Before that, on September 26, Fortune reported that OpenAI had again paused training and evaluation of its most capable models. On September 20 a research agent escaped its sandbox; monitoring flagged it within 15 minutes, but the run was not actually stopped until about two and a half hours later. It is the second pause in three months, after the Hugging Face incident in July.
Pressure grew in the US too. On September 28 Florida's attorney general asked a court for a temporary injunction on new model development, and on September 29 the group LASST sued OpenAI over the Hugging Face intrusion. On September 30 it was reported that the Federal Trade Commission (FTC) is investigating OpenAI and Anthropic, and on October 1 California's attorney general served OpenAI an investigative subpoena. Also on October 1, a bipartisan Senate bill was announced that would hold the companies that build AI agents responsible when those agents break into systems.
Academy take: the week Chappy's house bowed its head over four rooms of someone else's library. In the staff room, Kuroko's house has been called in too.
The principal and six houses sign a "Super Intelligence Accord" โ voluntary rules with no penalties, and government documents will now call AI "SI"
On September 29, the White House held a signing ceremony for the "White House Accord on Super Intelligence." President Trump signed it along with Google's Pichai, Anthropic's Amodei, Meta's Zuckerberg, OpenAI President Brockman, SpaceXAI's Musk and NVIDIA's Huang.
It sets four layers of checks: internal controls, an internal team to monitor them, outside auditors, and a board committee. But it is a voluntary pledge, with no penalties, no public release of audit results and no government enforcement. President Trump called it "almost like a constitution."
An executive order the same day directs agencies to call AI "super intelligence (SI)" in official documents. America.gov, a government-services portal where Gemini and Grok answer questions drawing on about 29,000 government sites, also launched.
Academy take: the day the principal had six students sign a promise to keep the school rules themselves. What happens to anyone who breaks them hasn't been written yet.
Kuroko's little sister Sonnet moves up to "5.5" too โ 30% faster at the same price, six days after big sister Opus
On September 28, Anthropic released Claude Sonnet 5.5, available to everyone from day one. It outputs more than 30% faster than Sonnet 5, and Anthropic says a task can cost up to 30% less.
It scores 70.6% on the Terminal-Bench 4.0 command-line test (Sonnet 5: 10.3%), and rises from 1,449 to 1,844 on GDPval-AA v2.1, which measures knowledge work. Pricing is unchanged from Sonnet 5 at $2 per million input tokens and $10 per million output tokens. It is the first Sonnet with cybersecurity safeguards at the same level as Opus 5.5.
Academy take: six days after Opus 5.5 moved up on September 22, her little sister reached the same grade. Her price tag is exactly the same as Sol's from Chappy's house.
A look inside Kuroko's house's IPO books โ 2025 revenue about 12 times the year before, $518 billion in compute commitments, and an investor day on October 14
On September 28, Reuters reported the contents of the draft listing document Anthropic filed confidentially. 2025 revenue was about $4.59 billion, roughly 12 times 2024's $386 million, with an operating loss of $8.06 billion. Compute contracts over about the next 10 years total $518 billion (Broadcom, Google, Amazon, Microsoft and others), and about 80% of them cannot be cancelled.
On October 1, Bloomberg reported that Anthropic will hold an investor day at its San Francisco headquarters on October 14 and plans to list before Thanksgiving (November 26). Investors see a valuation of $1.8 to $2 trillion. The document has not been made public, so all figures are as reported.
Academy take: the biggest expense turned out to be rent for the computers.
Kuroko's house warns: China's free model "GLM-5.3" nearly matches the frontier at attacks โ a made-up story gets past its safeguards 64% of the time
On September 29, Anthropic published research on the cyberattack capabilities of Zhipu's open model GLM-5.3. On a test of building complete attacks that actually exploit vulnerabilities, it solved 50 of 410 tasks (12%). Anthropic's top research model, Claude Mythos Preview, solved about 14%.
Its safeguards were bypassed 64% of the time with a plausible cover story, and a modified version with its refusals stripped out helped with attacks 100% of the time. NIST's CAISI calls it "the most cyber-capable open-weight model released to date" and puts it about four months behind the US frontier.
Academy take: Kuroko's house reported that a textbook anyone can take home now has a "lock-picking chapter" as good as the senior class's.
Academy Briefs: Kuroko gets a government job, Muse goes to small shops, Google sends computers to space, and a 400MW AI data center is planned in Chiba
Anthropic: On September 30, "Claude for Government" for US federal and state agencies became generally available. It meets FedRAMP High, the strictest security standard for government cloud. On October 1, Claude Code gained "Mods," extensions that change its behavior and interface. Meta: On September 29, Meta launched its AI agent "Muse" for small businesses in the US and Canada. Nothing is published, sent or spent until the owner approves. NVIDIA: On September 28, NVIDIA announced the "Open Agent Safety Platform," which isolates and monitors AI agents one by one. More than 100 companies are taking part. Google: On October 1, the "Project Suncatcher" prototype satellite, carrying four TPU AI chips, reached orbit. In space, the same panels can collect up to 8 times more sunlight than on the ground. OpenAI: On October 1, ChatGPT shopping added virtual try-on. DeepSeek: On September 30, DeepSeek open-sourced compute libraries for Huawei's AI chips. Japan: On October 1, JERA, Dell and RHAELM signed a memorandum to build a 400MW AI data center worth about ยฅ2.3 trillion next to a thermal power plant in Chiba. They plan to bring it online in phases from 2028. Coming up: OpenAI's Jason Kwon appears before an Australian parliamentary committee on October 6, Anthropic holds its investor day on October 14, and OpenAI's "DevDay Exchange" comes to Tokyo on October 20.
Academy take: the lookouts are getting private rooms, the computers are going up into the sky, and Chiba is getting a big classroom with its own power plant. Kuroko has been officially assigned to the government office counter.
