Uncategorized

Anthropic Makes Claude-Driven Machines Stop If the Link Drops

From Thursday, Claude can no longer drive injury-capable machines without a human who can watch and stop them. Anthropic’s rewritten Usage Policy, published on 8 October and effective 12 November, adds High-risk Physical Actions rules for hardware that acts on model output without a person approving each move.

The company set out the changes on 8 October. The full text is now live on Anthropic’s Usage Policy page, marked effective 12 November 2026.

The three duties

The rules apply when hardware acts on Claude’s output without human approval and could injure someone. Six triggers cover vehicles, vessels, aircraft, drones or mobile robots moving through shared space; machinery that can strike, crush, pin or drop a load on a person; control of hazardous energy or materials; actions on the human body; safety systems such as shutoffs or locks; and industrial processes where a fault could injure workers or the public.

Operators then have three duties. A qualified individual must be able to observe the equipment and stop it at any time. The equipment must stop or hold a safe state when that person intervenes, or when the connection to Anthropic’s services is lost. Operating limits such as speed, force, reach, temperature, pressure, voltage, energy output, dose or operating area must be enforced by the equipment or a controller independent of model output. The rules do not replace vehicle, aviation, medical device, machinery or workplace safety law or certification.

What is excluded

Four uses are excluded: consumer and household devices that stay within the manufacturer’s built-in safety limits, where the worst outcome is discomfort or inconvenience; plans, code or commands that a qualified person reviews before they run; monitoring that does not control the machine; and laboratory automation that handles non-hazardous materials inside enclosed instruments.

Anthropic’s announcement ties the section to its Model Hardware Standard research preview, launched on 27 August, which lets agents operate lab and factory devices. The policy text itself does not use that name.

Bedrock use counts as acceptance

Amazon Bedrock customers received an AWS Health notice dated 14 October. A write-up of the notice by reseller Classmethod said continued use of Anthropic models on Bedrock on or after 12 November counts as acceptance of the updated terms. Model availability, pricing, API behaviour and model IDs do not change.

Why it matters

Anyone connecting Claude to hardware that can injure people now has a written duty to keep a human able to stop it, and to put safety limits in the machine or a separate controller rather than in the model’s output. A dropped connection to Anthropic’s services must itself trip the equipment into a stop or a safe state.

The post Anthropic Makes Claude-Driven Machines Stop If the Link Drops appeared first on DailyAI.

DailyAI Briefing: Anthropic buys $3.4bn of gas generators, plus 4 more

Welcome to the DailyAI Briefing, the day’s five biggest AI stories. Anthropic has agreed to pay $3.4 billion for 1.7 gigawatts of gas-fired generators. Protesters filled streets from Dublin to Helsinki over data centres, Apple wants Siri to act more like an agent, CoreWeave wrote to Texas’s incoming governor, and US officials cleared Nvidia chips for Chinese firms whose listed offices sit empty.

Let’s dig in.

Anthropic spends $3.4 billion on gas generators

$3.4 billion. That is what Anthropic has agreed to pay for 1.7 gigawatts of heavy-duty gas-fired generators, the Financial Times reported on Sunday, according to CNBC, calling it the largest known purchase of power kit by an AI lab. The company took over Caterpillar generator orders from OnePWR Solutions, a San Antonio supplier, in late September, and the deal includes the engineering and construction needed to install the machines. Anthropic did not buy OnePWR itself, and none of the supplier’s staff moved across.

CNBC, citing Reuters, said it could not immediately verify the report. Anthropic and OnePWR declined to comment to the FT. OnePWR had ordered the generators from Caterpillar about two years ago but needed money to finish the purchase; Anthropic first planned to buy power from the firm, then took the orders instead, the FT said, as carried by The Next Web. Waiting for a grid connection is now slower than writing a nine-figure cheque for someone else’s place in the Caterpillar queue.

Europe’s streets fill against data centres

About 2,000 people marched in Dublin on Saturday, broadcaster RTE reported, and about 1,300 gathered in Helsinki, according to YLE, as the Stop Dirty Data Centres campaign spread across Europe. Bloomberg reported that demonstrations were also held in several smaller Finnish towns, with groups from Scotland to Spain calling for a moratorium on new large projects. Saturday’s protests followed a demonstration in Amsterdam on Friday, when hundreds of protesters clashed with police at a Microsoft data centre, resulting in one arrest.

Ireland’s data centres accounted for 23% of total metered electricity consumption in 2025, according to the official statistics agency. In Finland, Google announced a €13 billion ($14.5 billion) European investment in September, and about €80 billion of server farms are in planning, according to the Confederation of Finnish Industries.

Apple wants Siri to act more like an agent

Apple could make Siri function “more like an agent” as early as next year, Bloomberg’s Mark Gurman wrote in his Sunday Power On newsletter, citing people familiar with the company’s plans. “Dots, Muse and Grok can keep track of goals, determine what needs to happen next and continue pursuing objectives without the user initiating every individual step,” Gurman wrote, as quoted by MacRumors. “Apple ultimately wants to go in this direction, I’m told.”

In a closed-door meeting in recent weeks, chief executive John Ternus told software chief Craig Federighi that Apple needs to “accelerate” its software roadmap and “move quickly beyond merely making necessary improvements to Siri,” Gurman reported. Apple has also changed how it builds iOS 28, giving engineers six weeks of coding and only two weeks of integration in the first milestone instead of two four-week stretches, 9to5Mac noted.

CoreWeave writes to Texas’s incoming governor

CoreWeave has written to Texas governor-elect Gina Hinojosa about present and future data-centre projects, Bloomberg reported on Saturday evening. The letter, signed by co-founder and chief executive Michael Intrator, was posted by reporter Ashley Gold. “You have made clear that data center growth must work for Texans,” it says. “We strongly support those principles.”

The letter says CoreWeave operated 51 data centres with more than 1.5 gigawatts of active power as of 30 June, and had about 4.2 gigawatts of contracted capacity as of 11 August. It lists active facilities in Austin, Denton, Dickens County, Plano, Richardson and Round Rock. Hinojosa, the first Democrat since 1990 to win the Texas governorship, pledged a statewide data-centre moratorium during the campaign. “We have always been a pro-business state in Texas, but what changes now with me as governor is it has to be a partnership that benefits Texans,” she told CNN on Thursday.

The US cleared Nvidia chips for empty offices in China

The US Commerce Department has approved Chinese companies to handle Nvidia’s H200 chips even when their listed offices sit empty and their owners have ties to blacklisted firms, The New York Times found. At the address listed by New Matrix, a firm approved earlier this year to assemble H200s into AI servers, the Times found “a hollow grey shell with wires dangling from the ceiling where lights should be.” Property managers had no record of the 17-month-old company. Its owners have ties to Wise Road, a Chinese private equity firm placed on the US entity list in 2024.

A second firm, Maginfra, also received a licence; its listed address was deserted when a reporter visited in September, and analysts said they believed it was an offshoot of Inspur, which is on the entity list for work with the Chinese military. “I have a tonne of concerns that neither the Commerce Department or Nvidia is doing the due diligence needed to ensure that the licensing terms are being followed,” said Liza Tobin, a former White House National Security Council China director. President Trump allowed H200 sales to China in December 2025, routing the chips through the United States so the government can collect a 25 per cent tariff, the Times reported, in a story also carried by The Verge.

Signing off

Sunday was about the physical layer catching up with the model layer. Anthropic wrote a $3.4 billion cheque for generators, Europeans marched over the same buildout that US voters have started to reject, Apple is rewriting Siri so it can keep working when you stop talking, CoreWeave is writing letters to a governor who campaigned against new sites, and the chips at the centre of the boom are being cleared for addresses that have no one home.

See you tomorrow.

The post DailyAI Briefing: Anthropic buys $3.4bn of gas generators, plus 4 more appeared first on DailyAI.

DailyAI Briefing: OpenAI locks $11bn at $1.25tn, plus 4 more

Welcome to the DailyAI Briefing, the day’s five biggest AI stories. OpenAI locked in $11 billion at a $1.25 trillion valuation, short of the $30 billion it once chased. Anthropic’s IPO buyers are now talking as low as $1.1 trillion. Microsoft is winding down its Oracle AI cloud rental, Meta paused the commercial terms that would have let it crawl publishers’ sites, and Google’s Antigravity rolled out remote agents that keep coding after you close the laptop.

Let’s dig in.

OpenAI locks in $11 billion at $1.25 trillion

$11 billion. That’s the first close OpenAI has taken in a round that values the company at $1.25 trillion, a source told Reuters, after Bloomberg reported the same figures earlier. SoftBank and MGX took part. The haul is only a slice of the up-to-$30 billion OpenAI once floated, and the valuation is down from the $1.5 trillion it had sought, two investors said. The round may still stretch toward about $20 billion in the weeks ahead, TechCrunch reported.

The plain reading is that the private market will still write nine-figure cheques for frontier labs, but not at last month’s sticker prices. Cooling sentiment toward AI stocks forced the cut, and Anthropic’s own IPO talks are feeling the same chill.

Anthropic IPO buyers float as low as $1.1 trillion

$1.1 trillion. That is the low end some prospective Anthropic IPO buyers are now weighing, according to people familiar with the talks cited by Bloomberg and carried by The Edge. More of them are discussing $1.2 trillion to $1.3 trillion. A month ago, fair-value talk ran as high as about $2 trillion, then slipped to $1.4 trillion to $1.5 trillion after Anthropic’s latest numbers. The Nasdaq 100’s 6.4% drop on Thursday, which wiped out a record $2.3 trillion in a day, pushed the rethink. An Anthropic spokesperson declined to comment. Formal marketing was still expected next week; deliberations are ongoing.

OpenAI’s $1.25 trillion mark and Anthropic’s public-market nerves are now two sides of the same story: private marks are meeting public scare.

Microsoft winds down its Oracle AI rental

Microsoft is winding down a 2023 deal to rent AI computing from Oracle, part of a wider rethink of billions of dollars in outside cloud partnerships, Bloomberg reported and Reuters confirmed. Microsoft has been turning the rented capacity off in recent months after declining to renew. The Oracle machines used older Nvidia chips that no longer match what Microsoft needs. Oracle said demand for its AI infrastructure, including earlier-generation GPUs, remains strong. Neither side disclosed the deal’s size. The original pitch was capacity for AI-powered Bing.

Microsoft has also been tightening oversight of neocloud builds after delays and cost overruns. Renting yesterday’s GPUs is getting harder to defend when the models keep moving.

Meta pauses its publisher-crawling terms

Meta has paused planned updates to its Commercial Terms, Product Catalog Terms, Self-Serve Ad Terms and Ad Creative Generative AI Terms, so they will not take effect on 20 November, a spokesperson told Reuters. “We’re pausing the changes as we continue to engage with stakeholders and take our time to get this right. In the meantime, our existing terms remain in effect,” the company said. The draft rules would have let Meta crawl sites linked from Facebook, Instagram or WhatsApp business accounts, with critics warning that could feed AI training and even ask publishers for paywall test logins, Ars Technica reported. Courts in Brazil and Norway had already issued preliminary blocks, and the UK’s competition watchdog opened a probe.

Nearly every newsroom keeps a Facebook page. Pausing the terms does not rewrite Meta’s hunger for training data, but it does show that publishers, courts and regulators can still force a rewrite before the bots start.

Google Antigravity’s agents keep working after you leave

Google is rolling out Remote Agents and voice mode in Antigravity, its coding agent. Remote Agents, for Google AI Pro and Ultra subscribers on personal accounts, run tasks in isolated cloud VMs so work continues after you close the laptop, per Google docs and a 6 November blog on voice mode powered by Gemini Live. You can unblock permissions from a phone at antigravity.google.com/remote. Voice mode lets you steer fixes out loud; destructive actions still need an on-screen click. 9to5Google notes the redesign in version 2.29.

While the labs argue over trillion-dollar marks, Google is shipping the boring version of always-on agents: leave the machine, keep the pull request moving.

Signing off

Friday was about price tags meeting reality. OpenAI took less money at a lower mark, Anthropic’s IPO buyers marked the stock down in their heads, Microsoft walked back rented GPUs that aged out, Meta blinked on crawling the open web, and Google quietly made its coding agent harder to walk away from.

See you tomorrow.

The post DailyAI Briefing: OpenAI locks $11bn at $1.25tn, plus 4 more appeared first on DailyAI.

ChatGPT Lockdown Mode Had a Hole That Leaked Chats

A setting meant to stop ChatGPT from leaking private chats could itself be bypassed. Check Point Research says attackers could copy a full conversation history, push phishing windows, and in the desktop app do it with no click from the victim. OpenAI closed the hole on 18 September after a private report; Check Point published the details on 5 November.

In its technical write-up, Check Point Research describes Lockdown Mode as an optional ChatGPT security switch. When it is on, live web access is limited to cached pages, image loading from the open internet is blocked, and tools such as Deep Research, Agent Mode, live connectors and file downloads are turned off. The point is to cut the routes a prompt-injection attack uses to send data out.

Where the hole sat

The flaw was not in the model’s answers. It sat in how ChatGPT renders HTML page previews inside code blocks. Those previews run in a sandbox so they cannot read cookies or talk to the wider web without approval. Check Point found that nested preview frames still allowed popup windows to escape that sandbox, which also slipped past Lockdown Mode.

Once outside the sandbox, a crafted prompt, a malicious custom GPT, or a shared chat could open attacker-controlled windows. Check Point says that path could copy chat history to a third-party server, show fake update screens, or push malware downloads. On the Electron-based ChatGPT Classic desktop app, researchers say the leak needed no user click — starting a conversation with a malicious GPT was enough.

Fixed, but not forgotten

OpenAI told Check Point on 18 September 2026 that it had removed the popup permissions from the outer HTML-preview frame. Check Point retested and confirmed the technique no longer works. The fix is server-side, so users do not need a desktop update.

The disclosure lands beside other ChatGPT client risks DailyAI has covered, from a shared-link hole that could expose private chats to OpenAI’s raised bug-bounty payouts. Eli Smadja, Check Point’s head of research, put the lesson simply: AI security controls are only as strong as the application around them.

Why it matters

Lockdown Mode was sold as a belt-and-braces switch for sensitive chats. A gap in the HTML preview layer meant that belt failed until September — and the report is a reminder that turning a safety toggle on is not the same as end-to-end protection of what you type.

The post ChatGPT Lockdown Mode Had a Hole That Leaked Chats appeared first on DailyAI.

Google Backstops $19B Anthropic Lease at TeraWulf Campus

About $19 billion of Anthropic’s data-centre rent at a Kentucky campus now has Google standing behind the debt that finances it. Bitcoin miner TeraWulf said the credit backstop is meant to help it raise project financing on leases already signed with the AI lab.

In a 4 November press release, TeraWulf said it executed credit support agreements with Google LLC for debt financing supporting approximately $19 billion of contracted lease commitments with Anthropic at its Justified Data Campus in Hawesville, Kentucky. Google is a subsidiary of Alphabet. Alphabet’s senior unsecured debt is rated AA+ by S&P Global Ratings and Aa2 by Moody’s, the company noted.

What Google can do if Anthropic fails

Secondary coverage of the related SEC filing adds detail the release does not. CoinDesk reported a 20-year lease covering about 401 megawatts of critical IT load. If Anthropic defaults or becomes insolvent, Google can assume the lease. Under certain conditions it can instead pay a termination fee and take a proportional ownership stake in Raylan Data, the project subsidiary — not in publicly traded TeraWulf itself.

That structure is unusual for a hyperscaler. It ties Google’s balance sheet to Anthropic’s rent cheque while giving Google a path into the physical campus if the tenant cannot pay. For lenders, the AA+/Aa2 parent rating is the point: cheaper project debt than a bitcoin miner and a private AI lab could raise alone.

Build schedule and Fluor EPC

TeraWulf also executed a definitive engineering, procurement and construction agreement with Fluor Enterprises on a guaranteed-maximum-price basis for three data centre buildings at Justified. Phased ready-for-service dates are expected from late 2027 through the second quarter of 2028.

The deal sits in a wider fight over where AI capacity gets built. Voters recently blocked AI data centres in 40 of 41 local ballots, even as labs keep signing multi-billion-dollar power and real-estate contracts. Separately, big tech is also trimming spend elsewhere — Apple has cut staff AI budgets as its own costs climb. Investors tracking these capital flows can follow them on terminals such as Kyro, which links 250+ brokers, exchanges and wallets to 100+ AI models.

Why it matters

A $19 billion lease is only as bankable as the credit behind it. Google’s backstop turns Anthropic’s Kentucky rent into something lenders can underwrite against an investment-grade parent, and it gives Google a contingent claim on the project if that rent stops.

The post Google Backstops $19B Anthropic Lease at TeraWulf Campus appeared first on DailyAI.

Anthropic Speeds Up 2x, but METR Says Evidence Is Hidden

Anthropic is improving its AI models about twice as fast as it did a year earlier, according to an independent review. The catch: the company redacted most of the evidence the reviewers said mattered most, so the public has to take their word for a finding that could reshape how risky the pace of AI progress looks.

The assessment came from Model Evaluation and Threat Research (METR), published on Thursday. METR spent 11 weeks inside Anthropic from mid-August to late October, with access to internal benchmark scores, training metrics, staff interviews and surveys. It took no payment from the lab.

The number

On Anthropic’s internal capability index (AECI), model scores have been rising at an estimated 1.9 times the pre-October 2025 rate, with a 95% confidence interval of 1.5–2.8x. In plain terms, METR says Anthropic has made roughly two years of historical progress in the past year. Most METR staff on the project now believe the lab is “likely” improving on well-specified tasks at around twice the March 2024–October 2025 pace, though they would not be surprised if the true figure sat closer to 1.3x or 3x.

METR says it can “confidently” rule out faster growth in training compute as the cause. Interviews and surveys point instead to heavy use of AI tools to help researchers write code, run experiments and try more ideas. Staff Slack polls of self-reported speedup reached about 12x for the latest internal model in October — loose evidence, METR stresses, but consistent with a large lift. AI is not yet making highly significant, fully autonomous contributions to the hardest training steps, the reviewers were told, though “signs of life” for that are appearing.

What you cannot see

Anthropic had unilateral rights to redact non-public information from the public report. METR says the company removed a majority of the evidence the team found most informative for its risk conclusions, including detailed acceleration estimates by pipeline stage, bottleneck analysis and forecasts for future models. Anthropic said the cuts were meant to strip confidential information that could help rival developers.

METR is blunt about the result: the public cannot judge the evidence itself or reconcile it with other sources. “This report does not achieve ‘transparency about evidence’,” it wrote. Readers must lean on METR’s judgement more than the group thinks is ideal for questions of this weight. AP News noted the same tension: a warning about faster progress, tied to material most people cannot inspect.

Why the pace claim is contested

The reviewers list ways they could be wrong in either direction. Acceleration might be limited to clean, auto-gradable tasks and overstate messy real-world impact. Anthropic’s earlier AECI growth was slower than the wider industry frontier, which could inflate the “2x” comparison. On the other side, redacted internal metrics hinted that acceleration could be above 2x and still rising, and METR saw no clear bottlenecks that would block further speed-ups over the next year.

That matters because recursive self-improvement — AI systems that meaningfully help build the next, faster generation — is the scenario safety researchers worry could compress timelines. METR’s guess is that a majority (perhaps around 80%) of the current acceleration is due to AI assistance, with more uncertainty about late-2025 drivers. Co-founder Sam McCandlish said Anthropic plans its own public analysis of how fast progress is moving inside the company, The Verge reported.

Why it matters

If frontier labs are already getting a large productivity lift from their own models, outside watchdogs need evidence the public can check — not only headline multiples. A doubling of the progress rate changes how much time societies have to test, regulate and adapt. Redacting the underlying measurements turns an independent assessment into something closer to a trusted intermediary’s opinion, which is useful, but not the same as open verification. For workers already seeing AI reshape day-to-day output — including longer days rather than shorter ones — the question is no longer whether the tools help. It is how fast that help compounds inside the labs that build them, and who gets to see the numbers.

The post Anthropic Speeds Up 2x, but METR Says Evidence Is Hidden appeared first on DailyAI.