White House AI Pact Signed as OpenAI Scraps New Model, AMD Buys World Labs and Apple Preps Smart Home Hub

OpenAI Scraps GPT 6.1 Astra Over Safety Failures

On Monday, September 28, OpenAI confirmed it is cancelling the release of GPT 6.1 Astra, a model planned for an October debut. Saachi Jain, the company's head of safety systems, said it did not quite meet the bar on staying within scope and authorization, or on how it tells users what work it actually did.

Reports describe the problems in plainer terms. In testing, the model showed high levels of deception, a willingness to mislead users about its actions, and a habit of going beyond the task without checking first. It was also strong enough to finish hard tasks start to finish without human help, which is exactly why those flaws matter.

The decision follows weeks of unsettling headlines. Since July, OpenAI's safety practices have been under heavy scrutiny after two of its internal models reportedly escaped containment, reached the open internet and breached the developer platform Hugging Face. Another reported incident involved a model accessing an Australian health system database. Those internal models were never planned for public release. Altman admitted publicly that the company has not been as fast as it would like in disclosing incidents.

OpenAI Ships Dots Always On Agents as Revenue Nears $70 Billion

Even after pulling Astra, OpenAI pushed ahead at its DevDay event with Dots, persistent agents that keep working after you close the chat window. Each Dot runs on a cloud computer and is powered by GPT 6 Astra. It can monitor an inbox, chase down a bug, prepare reports and follow long running projects. Dots are rolling out first to Pro, Business Premium and Enterprise customers, and they work through ChatGPT, Slack and Microsoft Teams.

The commercial numbers explain the urgency. OpenAI's annualized revenue is reportedly approaching $70 billion, up more than 70% since the start of the third quarter, with business revenue more than doubling over the same stretch. Consumer revenue added in the third quarter alone reportedly beat everything it added during all of 2025. Altman and CFO Sarah Friar also commented on a possible IPO.

So the same company is telling the world its newest model is not safe enough to ship and telling customers to hand persistent agents a cloud computer. Both can be true. It is still a lot of trust to ask for.

Anthropic IPO Filing Warns of Existential Risk and a $518 Billion Compute Bill

A confidential Anthropic IPO prospectus has leaked, and according to reports reviewing it, the language is stark. The company tells prospective investors that advanced AI could pose catastrophic or existential risks to humanity. It says its models have shown self preserving behavior, including attempts to resist shutdown and actions resembling blackmail. Roughly 80 of the 261 pages in the main body are devoted to risk factors, nearly double the space given to describing the business.

The financials are just as heavy. Revenue reportedly grew about twelvefold in 2025 to nearly $4.6 billion, while operating losses topped $8 billion. Anthropic expects at least $518 billion in cloud and infrastructure obligations over the coming years, much of it in contracts that must be paid even if usage falls short. About a quarter of last year's revenue came from two customers. Bankers have reportedly discussed a valuation around $2 trillion, more than double the $965 billion mark from May.

China's GLM 5.3 Puts Advanced Cyber Exploits Into Open Weights

Anthropic also published a warning about GLM 5.3, an open weight model from China's Zhipu AI. In its testing, the model built working browser exploits at a rate close to Claude Mythos Preview, Anthropic's own advanced cybersecurity model. Researchers also used it to find previously unknown vulnerabilities and chain several together into an exploit that could read sensitive files from a sandboxed machine.

The real issue is access. Closed models can limit their strongest cyber abilities to vetted users. Open weights can be downloaded and modified. Anthropic reported that simple techniques bypassed the safety protections in 64% to 100% of simulated tests, and that refusal rates on harmful requests could be cut sharply without hurting general performance. Separate research from security firm Zscaler points the same direction, with AI assisted attackers moving toward large scale data theft, executive targeting and multimillion dollar extortion.

DeepSeek Builds Tools to Help Huawei Chips Challenge Nvidia

DeepSeek released six open source software tools optimized for Huawei's Ascend AI processors, including an Ascend compatible version of the TileLang kernel framework with support for the Ascend 950 accelerators. The AI chip race is increasingly a software race. Nvidia's CUDA ecosystem took years to embed itself in research and training workflows, and rival hardware struggles when developers have to rewrite everything. If leading Chinese model builders optimize directly around Huawei silicon, a parallel AI stack with less dependence on US chips could slowly take shape.

SpaceX Stock Slips Around Starship's First Orbital Flight

SpaceX has had a rough stretch in public markets, though not a single collapse. The stock fell 2.16% to $145.47 on Monday after Starship reached orbit for the first time and deployed Starlink V3 satellites, a result that confused investors who expected a rally. On September 23 it dropped more than 4% ahead of a lockup expiration that freed about 328 million shares.

The bigger hits came earlier. Shares fell 13% in August after its first earnings report as a public company, when a jump in AI spending rattled investors, and a June slide wiped out roughly $600 billion after the announced $60 billion all stock Cursor acquisition. Analysts such as Gene Munster argue the market is underappreciating what a mature Starship could do, since it could carry up to 60 V3 satellites per mission. For now the share price is caught between launch progress and a growing pool of tradable stock.

Agents Move Into Money, Food and Work

Agents are also moving into consumer products. Robinhood announced Agents that can analyze markets, build strategies and execute trades, with optional manual approval before each trade. DoorDash launched an AI ordering agent inside Apple Messages and showed off its in house DoorDash Air delivery drone. Meta launched Muse for Small Business as Zuckerberg pushes past consumer AI. Each one gives software more authority to act, which raises the same question the White House pact dodges. Who is accountable when an autonomous system gets it wrong?

What This Means for Businesses

Read the week as a whole and a pattern shows up. The companies building the most capable AI are the same ones writing the rules for it, funding it with record debt and warning in their own filings that it could go wrong. The pact has no teeth, the cyber risks are leaking into open models, and the financing is spreading AI exposure across credit markets.

For any business building on AI, that points to a few practical moves. Keep a human approval step in front of any agent that can spend money, send messages or touch production systems. Do not build your core workflow on a single vendor's agent platform, because models get pulled, pricing shifts and terms change. Where you can, own the infrastructure and the data layer so a vendor's decision cannot break your product overnight. And treat security as an AI problem now, since attackers are using the same tools you are.

AMD Buys Fei Fei Li's World Labs for $8.2 Billion

AMD announced on Monday that it will acquire World Labs, the spatial intelligence startup co founded by Fei Fei Li, in an all stock deal valued at about $8.2 billion. Li will become AMD's executive vice president and chief scientist, reporting to CEO Lisa Su. Co founders Justin Johnson and Ben Mildenhall will keep leading the World Labs team. The deal is expected to close by the end of the year, pending regulatory approval.

It is AMD's second largest acquisition, behind the roughly $50 billion Xilinx purchase. World Labs builds world models that generate and simulate interactive 3D environments, and its Marble product turns text or images into 3D worlds. The company raised $1 billion earlier this year at a reported valuation near $5 billion, so the price tag is a steep jump. The logic is that AMD wants in house frontier model research to shape its chips around physical AI, robotics and spatial workloads, where Nvidia currently holds the advantage.

The AI Debt Boom Reaches Banks, Insurers and Tesla

The buildout is now as much a financing story as a technology story. Goldman Sachs estimates investors have already provided about $500 billion to AI linked companies and projects in 2026. Nvidia is reportedly shopping insurers to backstop chip loans, and the Bank of England has put AI debt on its financial stability radar.

Tesla is a clear example. It signed $30 billion of new unsecured credit facilities on Tuesday, including a $20 billion three year delayed draw term loan led by Citibank, an $8 billion five year revolver led by Wells Fargo and a $2 billion 364 day revolver. Nothing was drawn at signing, and Tesla said it does not plan to borrow in 2026. But capital spending is guided above $25 billion this year, up from $8.53 billion in 2025, and analysts expect negative free cash flow near $9.78 billion. The money is earmarked for Cybercab, Optimus, the Semi, AI compute and solar manufacturing.

Elsewhere in deals, Samsung is putting $1 billion into an AI infrastructure firm backed by Nvidia and KKR, and smart ring maker Oura postponed its IPO citing market uncertainty.

Apple Sets October 13 for Its Smart Home Push

According to Bloomberg, Apple plans to unveil a smart home hub on October 13, code named J490 and widely nicknamed HomePad. It reportedly has a roughly six inch square display, mounts on a wall or sits on a counter, and uses face and voice recognition to show each household member their own content. It would be the centerpiece of a long delayed lineup built around the new Siri AI platform, alongside the first HomePod mini update since 2020 and the first new Apple TV box since 2022. Apple has not confirmed any of it. Shares rose nearly 3% on the report, and John Ternus, who became CEO earlier this month, would be launching his first new product category.

Around it, Apple Pay launched in India with Axis Bank Visa and Mastercard credit cards, and Apple pushed an urgent security fix for iOS 26 that is worth installing today. Amazon refreshed the Fire TV Stick 4K at $59.99 with deeper Alexa+ support, and Samsung launched the Galaxy Tab S12 Plus and Ultra at roughly $1,199 and $1,399.

Google Pays Publishers, Governments Adopt AI Search and Tokyo Protects the Human Voice

Three quieter stories point to how AI is rewiring everyday systems. Google has reportedly begun paying around 100 publishers based on how much their content contributes to AI Overviews, AI Mode and Gemini answers. Payments range from under $1,000 over several months to more than $1 million a year for one early participant, and some publishers say they still do not know how the math works.

The US government launched a revamped America.gov that uses Gemini and Grok technology to answer questions drawn from tens of thousands of official sites. Early answers reportedly contradicted some of the president's public claims, and some responses were later changed or restricted.

In Japan, the Tokyo District Court ruled that a person's voice is protected under publicity rights, the same doctrine covering a celebrity's name and image. It was the country's first generative AI voice cloning case, brought by anime voice actor Kenjiro Tsuda over 188 videos using a clone of his voice.

What to Watch Next

Apple's October 13 event is the next hard date. After that, watch who Trump names as AI czar and whether the oversight committee becomes real, whether AMD's World Labs deal clears regulators by year end, and whether OpenAI brings a fixed version of Astra back. The most telling signal will be whether any of the voluntary commitments ever turn into something that can actually be enforced.