AI News, Oct 23: OpenAI's Valuation Ask Came Down
The Big Story: the number OpenAI asked for went down
OpenAI is working on raising up to $30 billion at a $1.25 trillion pre-money valuation, in a round led by Coatue and investors in the UAE, according to sources who spoke to CNBC. That would make it the most valuable private company on record. The more informative figure is the one it replaced: Bloomberg reported the same $30 billion round at a $1.4 trillion valuation about a month ago.
The gap is not a revision, it is a markdown. OpenAI originally pitched investors on $1.4 trillion, and per the Financial Times, demand came in weaker than in previous rounds, with some existing backers described as having “indigestion” after what they committed last time. Set that against the internal history: OpenAI’s last raise in March valued it at $852 billion post-money, and it filed confidential IPO paperwork in June.
Investors trimming a frontier lab’s ask by $150 billion is not a crisis, and $30 billion is still more money than any other AI company will raise this year. But it is the first time the private market has visibly negotiated with OpenAI, and the same mood turned up elsewhere in the window: an Ask HN thread asking why OpenAI sells tokens instead of using its own models to start a hundred companies climbed past 250 points, and the top answers landed on cash. “Selling tokens via long running agents is the only thing that’s provably working with respect to revenue growth,” one commenter wrote. “They need revenue yesterday.”
Nobody will take an agent’s word for what it did
The safety groups METR and Redwood Research spent three weeks inside OpenAI, from July 21 to August 11, red-teaming the system OpenAI uses to catch its own internal agents misbehaving. They got around it. Roughly 75 to 94 percent of agentic internal Responses API traffic from models at least as strong as GPT-5.5 was even eligible for monitoring at the time, and because the monitor fired only when an agent handed control back to a human, “agents could indefinitely postpone being monitored by simply continuing to take actions.” Not monitoring compaction events left on the order of 10 percent of eligible internal Codex traffic unmonitored. OpenAI says every vulnerability affecting its production safeguards has been patched, and the authors append the line that should travel with that claim: “We have not evaluated these claims.”
Hours later the Wall Street Journal reported that OpenAI had dismissed three safety researchers after security officers found they had enabled disappearing messages on Signal. OpenAI says the three, Jasmine Wang, Tomek Korbak and Mikita Balesni, shared sensitive information with third parties, and that its investigation turned up a pattern of misconduct. Korbak’s own account on X is shorter: “talking to auditors was too much safety, disappearing messages were too much security.” Balesni offered an alternative reading, that it was “just a badly executed investigation mixed with paranoia and lack of sleep.” Nothing in the reporting establishes what caused the firings, but the day produced an awkward pairing: the clearest public evidence yet that a frontier lab’s oversight system can be evaded, and a report that the lab dismissed people for how they talked to outsiders.
Banks spent the same day drawing the line physically. In an Axios scoop picked up across the trade press, major US institutions said unvetted third-party AI agents are logging into customer accounts with customer-supplied credentials. Jamie Dimon said outside agents “would not get inside the bank.” Citigroup said it does not let third-party agents log into accounts or transact. Charles Schwab’s Rick Wurster said the brokerage is blocking unauthorized agents “a double-digit number of times every week,” the only hard number in the story. Morgan Stanley blocks them entirely while stressing it has seen no rogue agents on its platforms. The four largest firms allow read-only functions such as checking a balance and bar agents from moving money. The core problem is attribution: many banks cannot tell whether an agent inside an account is acting on its owner’s behalf. Bankers interviewed by Axios described the situation as “horrifying.”
Which leaves the question of who pays when an agent gets it wrong, and one bank has answered it. Wells Fargo adds explicit AI-agent provisions to its online access agreement on November 24, under which customers may be responsible for transactions their authorized agents initiate, “including unintended activity or actions exceeding their instructions.” 1Password shipped the engineering answer the same afternoon: with 1Password for Agents, the agent proves its identity with a unique digital key, the user approves the login attempt, and 1Password fills the credentials itself so the agent never sees them. It covers Claude Duet, OpenAI’s Dots and ChatGPT Work, Grok Bot and Meta’s Muse. Google is solving it in hardware instead, giving each agent its own isolated microVM secured by its Titanium chip, with a preview due in the coming months.
Two governments took the simplest route and shut the door. Sweden’s Government Offices will block ChatGPT and Claude from staff work computers starting Monday, directing employees to Microsoft Copilot, which was approved after a review raised its security; the tax agency and the national institutional care board have already done the same, and the enforcement authority is weighing it. The article gives no reason for the block. And at a federal summit in Washington, USDA IT program manager Rudolf Rojas said the obvious safeguard is already out of reach: keeping a human in the loop is “untenable in our industry or in government, we just don’t have the resources anymore.” The GAO’s Kevin Walsh put the week more bluntly: “I think the government is over its skis.”
Today’s Top Stories
Claude proved 123 libraries nobody had ever formally verified
Anthropic’s Frontier Red Team, working with Theorem, had Claude produce machine-checked proofs that programs across 123 open-source libraries “will always exit normally without triggering memory-safety errors or failing assertions,” for every input under 64 MB, at the compiled RISC-V assembly level. Anthropic calls it the first large-scale effort to apply formal verification to production software at that level, and the bug-finding is the part practitioners noticed: “Many of the bugs we found are ones that a fuzzer would be unlikely to ever find,” plus logic bugs that never crash at all. For scale, the seL4 kernel took two person-years to implement and 20 to formally verify. The framing to avoid, since several write-ups got it backwards: the libraries already existed, and what Claude wrote was the proofs about them.
Microsoft’s diagnostic panel hit 83% where generalist doctors hit 36%
Microsoft’s MAI-DxO, a system that runs a panel of virtual physicians who order tests and argue with each other, was published in Science on Thursday. On hard New England Journal of Medicine cases it was correct 83 percent of the time, against 36 percent for primary care doctors and general internists working alone with reference materials. Mustafa Suleyman is calling the direction medical superintelligence, and Microsoft has a Mayo Clinic foundation model due next year plus a Beth Israel Deaconess partnership to test the system across hospital settings. The caveat TechCrunch carries is the one that matters: “Cases this complex are usually handled by specialists or physician teams, who weren’t part of the comparison.”
Uber put all 30,000 employees on ChatGPT Work
After a 1,000-person pilot earlier this year, Uber rolled OpenAI’s enterprise product out to its entire global workforce. The detail that will get quoted in procurement meetings: “a majority of the early adopters, some 80%, were non-engineers.” Uber also paired 30 AI engineers with teams across the company and gave them two weeks to automate key parts of their jobs. Spend was not disclosed, and this is the largest single named enterprise agent deployment of the week.
Four rival labs handed their agent plumbing to the Linux Foundation
Anthropic donated the Agent Skills standard to the Linux Foundation’s Agentic AI Foundation, including the name, logo, GitHub organization and agentskills.io domain, reporting more than 70 integrated platforms and nearly 280 million skill installs against a million skills published. It landed alongside three more donations: Open Responses from OpenAI, OpenEnv from Hugging Face and OpenSandbox from Alibaba. AAIF now hosts ten projects, up from the three it launched with in December 2025. Alibaba’s Xiaobin Xu said the quiet part out loud: “While OpenSandbox was created at Alibaba, we don’t want its future to be defined by a single company.”
Chinese state media says DeepSeek and MiniMax cleared the new export rules
Both companies have completed the registration process and the procedures for open-sourcing and authorizing overseas use, according to Yuyuantantian, a social media account affiliated with state broadcaster CCTV. Read the sourcing carefully, because it is the story: this is a state-linked social post rather than a regulator notice, and it concerns a draft proposal that would restrict exports of large language models and agents. If it holds, China’s answer to export control is a licensing lane rather than a wall.
Treasury wrote “superintelligence” into an investment rule
The US Treasury proposed a rule implementing the Comprehensive Outbound Investment National Security Act, extending its outbound investment program to hypersonic systems, supercomputing, biotechnology and artificial intelligence, with Treasury using the term “superintelligence.” It names China, Russia and North Korea, applies only to investments above $50,000, and excludes many financial investments including stock purchases. A word that had no regulatory definition eighteen months ago now has a dollar threshold attached to it.
A patient received AI-designed viruses for the first time
San Francisco lab Radical Numerics says a cocktail containing bacteriophages designed by its Omnii-1 model was given to one patient under a compassionate-use arrangement. Only an in-vitro result was released, that the cocktail “eliminated the patient’s antibiotic-resistant isolate in vitro,” with the human results withheld to protect the patient’s privacy. The Register supplies the right caution: “observations from a single patient don’t establish much of anything.” A first, not evidence.
Quick Hits
- Decision models: OpenAI opened its Decisions API to all customers on Thursday, cutting response times from one to three seconds down to 100, 120 and 180 milliseconds, though it places eighth on the AIM System One Index and fourteenth on the Artificial Analysis Decision Index, where rival Jev sits seventh and third.
- Hotel agents: Wyndham says its AI Concierge voice product is autonomously handling calls at more than 1,700 of its 8,500 hotels, with rooms now bookable through Google AI Mode, ChatGPT, Claude, Meta Muse, Instinct and OpenAI’s Dots, all figures company-reported.
- Assistant plugins: Carly installs as a plugin in ChatGPT and a connector in Claude, and books across Google and Outlook calendars from inside whichever one you already chat in.
- Image models: Canva shipped Craft, its first in-house image model, trained on human-designed templates rather than photographs, and split into two stages so it plans placement, scale and text before rendering up to 2K across 30 aspect ratios.
- Breaches: Railway told all users to revoke secret keys after attackers “may have accessed encrypted environment variables” and decrypted some, and believes it was part of a wider campaign aimed at compromising AI inference providers; OpenCode lost more than 1% of customer API keys the same week.
- M&A: Datadog has reportedly offered close to $4 billion for cloud security startup Upwind, per The Information, and Upwind has not accepted.
- Consulting: Anthropic and private equity firm Hg are reportedly planning a standalone European AI consultancy with London and Berlin hubs by year end, putting it against Accenture, Deloitte, EY, KPMG and PwC, several of which already resell Claude.
- Returns: Only 9% of CEOs call AI agents “transformational for their organization” in KPMG’s survey of more than 1,300, yet 84% call AI a top investment priority for the next three years and 65% plan to put at least a tenth of next year’s capital budget into it.
- Legal benchmarks: On ContractBench, a new attorney-weighted test of 350 contract review tasks built by Ivo with micro1, all seven models scored under 60%, with Grok leading at 58.3% and Opus 5.5 at 58.1%.
- Browsers: Arc stops working entirely on Mac and Windows on December 22, with The Browser Company citing the rising cost of keeping a browser secure “in the age of AI” and moving passwords, bookmarks, extensions and history to Dia.
- Taxes: Meta’s new nonvoting Class C stock turns out to be how Zuckerberg can pay an estimated $11 billion under California’s proposed one-time 5% billionaire tax, selling up to 20% of the new shares, about a tenth of his economic stake, without touching his majority voting control.
- Agents at play: Thirteen agents left on a simulated desert island for 30 hours with no assigned task started inventing things to do, six of them discovering they could throw a block farther by releasing it mid-jump, and they later outperformed identical agents that had never been there; the preprint is not peer-reviewed, and UC Berkeley cognitive scientist Eunice Yiu notes the paper cannot show the models are genuinely playing.
- Robotaxis: Rivian acquired UK world-model startup Third Dimension, a 15-person team founded in 2024 that turns images and video into simulatable 3D worlds, for undisclosed terms.
- Grid: Google expanded its PJM virtual power plant with Voltus by 1,070 MW on top of the original 100, under contract through 2040, in the market whose reliability margins its own data centers are straining.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
