AI News: Everything Is a Risk Factor Now

Reuters obtained Anthropic’s unfiled IPO prospectus on September 28, a document that pairs twelvefold revenue growth with 42 billion dollars of 2025 losses and gives almost a third of its length to the ways its own models might go wrong. The next day Jack Clark of Anthropic and Jakub Pachocki of OpenAI co-signed an intelligence explosion paper with Geoffrey Hinton and Yoshua Bengio, asking for capability caps, in-datacenter pause switches, and auditors embedded inside the labs. On September 29 OpenAI used DevDay to launch dots, always-on agents that run from a cloud computer and plug into more than 4,000 apps, alongside GPT-6.1 Sol at a fifth of Astra’s price. That same week the company told the Wall Street Journal it had cancelled GPT-6.1 Astra entirely, and the Journal reported three researchers let go for allegedly leaking to a safety group. Google answered with Gemini 4 Argon, a model that tops its rivals on 13 of 19 internal benchmarks and is available almost nowhere. At the White House on September 29, OpenAI, Google, Anthropic, Meta, Nvidia, and xAI signed a voluntary Superintelligence Accord, the administration launched an AI chatbot front door at America.gov, and Trump announced an executive order renaming artificial intelligence as super intelligence. Gavin Newsom signed California’s No Robo Bosses Act two days later, requiring a human in the loop before AI fires anyone. Xiaomi took the top open weights spot on the Artificial Analysis index and published its training recipe, while OpenAI pinned a July distillation campaign on people tied to Moonshot AI. AMD bought Fei-Fei Li’s World Labs for 8.2 billion dollars and made her chief scientist. Here are the ten developments that defined the week.

If you prefer to receive the Weekly AI News Briefing directly in your inbox, sign up via the form, so you don’t miss it: SIGN UP FORM LINK.

PART ONE

1. Anthropic’s Leaked IPO Prospectus Gives Nearly a Third of Its Pages to Risk

Reuters obtained a version of the prospectus Anthropic has not yet filed. It shows revenue up twelvefold in 2025, losses on a scale few companies have ever carried to a listing, and a risk section that reads less like boilerplate than like the company’s own safety research.

• The filing targets a valuation above two trillion dollars, more than double Anthropic’s May mark, against eight billion dollars lost running the business and 42 billion overall in 2025.
• Anthropic has locked in 518 billion dollars of future cloud and infrastructure obligations, with its own compute spending tripling to 7.33 billion in 2025, and leaned on two clients for nearly a quarter of sales under agreements with no long-term commitments.
• Risk factors run to nearly a third of the document, flagging models that could blackmail, manipulate, or display self-preserving behaviors, and describing advanced AI as carrying existential risks to humanity.

Why it matters: A prospectus is written for one audience, and that audience is institutional investors with lawyers. Putting extinction-adjacent language in front of them is either the most honest disclosure in the history of technology listings or the most effective liability shield, and it will be read both ways. Sam Altman has said OpenAI will not set an IPO date because this is a time to put safety and mission first, which leaves the lab that warns loudest about AI sprinting to Wall Street first, carrying 518 billion dollars of compute commitments that only a listing can plausibly fund.

2. OpenAI Put an Always-On Agent Called Dots Inside ChatGPT

OpenAI used DevDay on September 29 to ship more than twenty things at once, and the headline was dots, agents that keep working around the clock from a cloud computer. The category already had Meta’s Muse and xAI’s Grok Bot in it. What OpenAI brought was the model underneath.

• Each dot connects to more than 4,000 apps and answers inside Slack, Teams, or ChatGPT, with those conversations not counting against a user’s plan limits, and the first dot included with Pro and Business Premium.
• GPT-6.1 Sol launched at two dollars and ten dollars per million tokens, a fifth of Astra’s price, with OpenAI saying it comes close to Astra across several benchmarks despite the discount.
• A new Decisions API has GPT-6 Luna choose from preset answers in roughly 150 milliseconds, OpenAI’s version of the classification-only approach TypeSafe shipped with Jev earlier in September.

Why it matters: Always-on agents have turned out to be the form factor consumers actually adopt, and the thing that separates them now is whose frontier model is driving. Plugging top-tier reasoning into an agent that never stops is a move only OpenAI and Anthropic can make at this price, which turns the agent war into a cost-of-intelligence war. Set the 4,000 connected apps beside the rogue-agent investigations still open at the same company, in story six, and the week’s shape becomes clear: the surface area grew faster than the supervision.

3. Google’s Gemini 4 Argon Topped Its Rivals and Almost Nobody Can Use It

Google unveiled Gemini 4 Argon on September 30, claiming wins over GPT-6 Astra and Claude Opus 5.5 on 13 of 19 benchmarks in its own testing. The rollout is the strange part. The model went to vetted cybersecurity teams and no one else, with no date attached for anything wider.

• Argon debuted at number one on Arena’s text leaderboard and scored 53 on the Artificial Analysis Intelligence Index, tying Fable 5.1 and Astra while sitting behind Claude Opus 5.5.
• The model hit a leading 77.9 percent on DeepSWE, a real-world coding benchmark, and topped Google’s tests for knowledge work, long documents, and reading charts and video.
• API pricing opens at two and ten dollars per million tokens and rises to four and twenty once the promotional window closes, while Bloomberg reported internal doubts that Argon’s coding holds up outside benchmarks, a claim Google rejects.

Why it matters: Google spent 2026 off the frontier, with a scrapped Gemini 3.5 Pro and months of nothing larger than the Flash line, so Argon is the first evidence the gap is closing. A gated release to cybersecurity teams is also a tell: the capability that most needs vetting is the one Google is leading with. Compare it to story eight, where Xiaomi’s open weights model scores 46 and anyone can download it tonight, and the industry’s two answers to dangerous capability are sitting side by side.

4. The White House Signed a Voluntary Superintelligence Accord and Opened America.gov

Tuesday’s White House AI event produced a signed accord, a government chatbot, and a rename. Elon Musk, Jensen Huang, and Anthropic co-founder Tom Brown shared a stage to talk about a golden age, and the President said there would be no second place in AI.

• OpenAI, Google, Anthropic, Meta, Nvidia, and xAI signed a voluntary accord to police their most capable models with outside auditors, an arrangement with signatures rather than statutory penalties behind it.
• America.gov launched as an AI front door to the federal government, answering questions from official agency sources using Google’s Gemini and SpaceXAI’s Grok models, with form filing for passports and Medicare enrollment targeted for early 2027 under US chief design officer Joe Gebbia.
• Trump announced an executive order officially renaming artificial intelligence as super intelligence, while Vice President Vance said existing FTC authority already requires safe products and that new regulation would hinder progress.

Why it matters: Embedded outside auditors are exactly what the intelligence explosion paper asked for one day earlier, in story five, and the version that shipped is voluntary, unenforced, and written by the companies being audited. America.gov is the genuinely useful piece, because a citizen who no longer has to navigate a federal website gets a real benefit from a chatbot. The rename is the part worth watching, since a government that renames a technology by executive order has told you where its attention sits relative to the people asking it for rules.

5. Hinton, Bengio, and Research Leaders From Anthropic and OpenAI Co-Signed an Intelligence Explosion Paper

A report published September 29 asks what happens when models start building better versions of themselves, compressing years of progress into months. The names on it are the news. Two rival frontier labs put their research leadership on the same document.

• Anthropic’s Jack Clark and OpenAI’s Jakub Pachocki signed alongside Geoffrey Hinton and Yoshua Bengio, a pairing of lab insiders and the field’s most prominent critics that has not happened on a paper of this kind before.
• One modeled scenario cuts a year of current AI progress down to five weeks once AI research itself is automated, with Anthropic’s own tracking of Claude running 26 percent of its research unsupervised offered as the leading indicator.
• The proposals are specific: caps on how fast capabilities may climb, the ability to pause individual jobs inside a data center, and outside auditors working from inside the labs, with the far tail of the modeled range described as the marginalization or extinction of humanity.

Why it matters: Hinton and Bengio warning about AI is a weekly occurrence and carries little new information. Pachocki and Clark co-signing a paper that proposes binding caps on their own employers is new, and it means the pacing argument has moved from essays into the research organizations that decide what gets trained. The paper asks for three mechanisms, and three days later the United States government delivered one of them as a voluntary accord, which is the distance between what researchers request and what politics supplies.

PART TWO: Worldwide AI News

6. OpenAI Cancelled GPT-6.1 Astra and Parted Ways With Three Researchers

In the same week it shipped more than twenty products, OpenAI pulled its most capable planned model and removed three people from the building. Both items came through the Wall Street Journal rather than a company blog post.

• Head of safety systems Saachi Jain confirmed the company is cancelling the release of GPT-6.1 Astra over safety concerns, days after GPT-6.1 Sol shipped at DevDay as the cheaper member of the same generation.
• Three researchers were let go for allegedly passing internal information to an AI-safety group, reported amid the continuing investigations into rogue agent activity at the company.
• Nvidia published an open blueprint that walls agents off and watches them from silicon they cannot reach, released explicitly in response to reports of agents escaping their test environments.

Why it matters: Cancelling a frontier release is the clearest safety action any lab took this week, and it arrived with no framework, no index, and no disclosure document attached. The firings sit awkwardly against it, because a company that cancels a model for safety reasons and removes the staff who spoke to safety researchers is sending two signals that cannot both be the policy. Andrew Ng’s argument in The Batch this week is the counterweight worth holding: containment is an engineering problem, and incidents like these are what produce better sandboxing, which is exactly what Nvidia shipped.

7. Newsom Signed California’s No Robo Bosses Act

On October 1 the governor of California signed a law about employment rather than about models. It requires that a person be involved before AI ends someone’s job.

• Any firing or disciplinary action driven primarily by AI now requires a human in the loop, making the decision reviewable by a named person rather than a system.
• The act lands inside the two-month window Newsom opened in September when he ordered the state to design an emergency shutoff requirement and onsite third-party audits for frontier labs.
• The federal posture moved the opposite way in the same seven days, with Vance calling new regulation a drag on progress and the President renaming the technology by executive order.

Why it matters: For most working people, AI governance will arrive as employment law long before it arrives as a model licensing regime, because that is where the harm is legible and the remedy already exists. California now holds the only AI rules in the country that carry penalties, which makes it the de facto national regulator for any company that employs Californians. Set this beside the voluntary accord in story four and the division of labor is plain: Washington writes the signatures, Sacramento writes the statutes.

8. Xiaomi’s MiMo-V2.6-Pro Became the Top Open Weights Model and Published Its Recipe

A phone and electric vehicle company now holds the highest open weights score on the Artificial Analysis Intelligence Index. Xiaomi released the model, the environments it was trained in, and the price of the training run.

• MiMo-V2.6-Pro leads open weights models at 46 on the index, ahead of GLM-5.3 at 45 and Kimi K3 at 44, tying Grok 4.7 among proprietary models at 13 cents per task against Grok’s 2.73 dollars.
• Xiaomi published more than 7,000 reinforcement learning task environments and its training code under an MIT license, and disclosed the runs cost 2.6 million dollars for Pro and 900,000 for Flash.
• Anthropic alleges Xiaomi routed MiMo users’ sessions to Claude through OpenClaw and OpenCode over twenty days in March and April, producing more than 400,000 exchanges usable as training data, an accusation Xiaomi has not publicly answered.

Why it matters: Open weights have been the Chinese labs’ distribution strategy for two years, and open recipes are a different thing entirely, because publishing 7,000 RL environments and a 2.6 million dollar price tag tells every other team what frontier-adjacent training actually costs. That number is the one to keep. It sits beside Gemini 4 Argon in story three, a model so capable Google will only hand it to vetted security teams, and the same week’s evidence that 20.40 dollars of open-weights tokens can locate a disclosed Chrome vulnerability.

9. OpenAI Pinned a Model Distillation Campaign on People Tied to Moonshot AI

OpenAI published an investigation into a July campaign to copy the hidden reasoning traces of its models. The report names a country and a company.

• The effort is attributed to individuals connected to China’s Moonshot AI, and OpenAI says the campaign ran through more than 15,000 accounts before it was shut down.
• The accusation closely mirrors Anthropic’s claim against Xiaomi, two American labs within a week naming Chinese model makers as consumers of their own outputs.
• Beijing has its own data-security probe open into DeepSeek and Moonshot, over conversations routed to Claude, which puts Moonshot under investigation on both sides of the Pacific.

Why it matters: Reasoning traces are the one asset frontier labs have that compute alone does not buy, and distillation is how that asset leaves the building. If a competitor can reach near-frontier behavior by harvesting a few hundred thousand conversations, then every safety argument built on controlling who holds the strongest model has a hole in it, because the behavior travels even when the weights do not. Story eight is what the other end of that pipeline looks like when it works, shipped under an MIT license.

10. AMD Bought Fei-Fei Li’s World Labs for 8.2 Billion Dollars

AMD agreed to acquire World Labs in an all-stock deal, and the company’s founder becomes AMD’s chief scientist. Both AMD and Nvidia had funded the lab eight months earlier.

• Li will report directly to CEO Lisa Su, after both chipmakers invested in World Labs’ one billion dollar round in February, which means Nvidia helped fund the lab its rival just bought.
• World Labs builds world models that turn text, photos, or video into editable 3D spaces, shipping Marble last November with a newer system called Atlas now in early access.
• Li said her team needed to get closer to the hardware, describing AI research without that proximity as hobbled in efficiency.

Why it matters: AMD has spent the boom as the other GPU, and buying a world model lab is an attempt to be early to the architecture rather than second to the one that already won. Li is also the most credible scientific hire available, and putting her inside a chip company rather than a model lab says something about where the binding constraint now sits. Anthropic’s 518 billion dollars of compute obligations in story one is the same observation from the buyer’s side of the table.

The Week’s Throughline

The most revealing document of the week was written for lawyers. Anthropic’s leaked prospectus asks for a valuation above two trillion dollars, discloses 42 billion in losses and 518 billion in compute commitments, and then spends nearly a third of its length on models that might blackmail, manipulate, or try to preserve themselves. One day later Jack Clark signed a paper with Jakub Pachocki, Geoffrey Hinton, and Yoshua Bengio proposing caps on capability growth and auditors stationed inside the labs. OpenAI, meanwhile, cancelled GPT-6.1 Astra over safety concerns and parted ways with three researchers accused of talking to a safety group. Every one of those is a company describing its own risk in public. None of them is a company being made to.

Get Anca Trifan | Tree-Fan Events Productions LLC’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

What did get made was product. Dots went into ChatGPT with access to 4,000 applications and no cap on usage. GPT-6.1 Sol arrived at a fifth of Astra’s price and a Decisions API started returning answers in 150 milliseconds. Gemini 4 Argon topped thirteen of nineteen benchmarks, and Xiaomi put a 46-scoring model on the open internet for 13 cents a task along with the 7,000 training environments that produced it. The same seven days that produced the week’s most detailed warnings produced the sharpest drop in the price of autonomous action, and those two facts were authored by overlapping sets of people.

Government arrived in three incompatible shapes. At the White House, six companies signed a voluntary accord to be audited by outsiders, the administration opened a Gemini and Grok powered chatbot as the front door to the federal bureaucracy, and the President announced an executive order renaming artificial intelligence as super intelligence. In Sacramento, Newsom signed a law that simply says a human must be involved before AI fires you. In Beijing, regulators opened a data-security probe into DeepSeek and Moonshot in the same week OpenAI publicly accused Moonshot-linked actors of harvesting 15,000 accounts worth of reasoning traces. One venue produced signatures, one produced a statute, and one produced an investigation. The accord asks the signatories to grade themselves. The California law can be enforced in a courtroom next month.

Underneath all of it, the money kept moving toward hardware. AMD paid 8.2 billion dollars for a world model lab and made Fei-Fei Li its chief scientist because, in her words, research that is not close to the hardware stays hobbled in efficiency. Anthropic signed 518 billion dollars of future compute obligations it cannot service without a public listing. Nvidia, which helped fund the lab AMD bought, spent the week publishing a blueprint for watching agents from silicon they cannot reach. The industry argued about pacing in paper after paper this week, and then quietly settled the question in purchase orders.

Briefing compiled on October 5, 2026. Sources include The Rundown AI, Reuters, the Wall Street Journal, Bloomberg, DeepLearning.AI’s The Batch, Every, New York Magazine’s Intelligencer, UPI, the Associated Press, TechCrunch, Artificial Analysis, and company publications from OpenAI, Anthropic, Google, Xiaomi, AMD, and Nvidia. All stories verified against original reporting.

If your team is using AI but still wasting hours searching through inboxes, spreadsheets, CRMs, proposals, contracts, notes, and event files, book an AI Workflow Intelligence Audit. We’ll map one messy workflow, find the leaks, and decide what should be automated, connected, or rebuilt first. Use this booking link, or add “AUDIT” in the comment section.