AI Policy · Daily

Anthropic told investors its next models may soon cross a safety threshold for AI that sharply speeds up the company's own research, though it says no current model has done so yet. Britain's cyber agency expects AI-enabled malware in ransomware attacks on critical infrastructure by 2028, saying Russia and China will start using AI routinely to scout targets and write malicious code. OpenAI will let utilities throttle power to future U.S. data centers it directly develops during extreme weather and emergencies, starting with a planned Georgia campus, and will pay for the grid connections the site needs. Seven popular chatbots gave outdated or inaccurate answers to voting questions in about 14% of October tests, with Anthropic's Claude and Microsoft's Copilot accounting for three-quarters of the clear errors.

I.Top Stories

Anthropic IPO filing says it may soon cross own threshold for AI research speedup

Anthropic's amended IPO registration says internal estimates for future models, not current ones, indicate it "may imminently meet" the threshold for "dramatically accelerated progress attributable to AI" set in its Responsible Scaling Policy, the internal rulebook that tightens safeguards as models improve, TechCrunch reported. The filing says no current model has been designated as exceeding the threshold, and that Anthropic has in the past failed to spot a significant speedup in its development until after it happened. Anthropic also says it is weighing mandatory human sign-off on some research tasks its models now run automatically. Sen. Elizabeth Warren (D-Mass.) set today as the deadline for Anthropic to answer her questions on risks in its IPO paperwork, as reported by The Wall Street Journal in AIPD's October 28th edition.

Read at TechCrunch ↗ • Read at SEC ↗

UK cyber agency expects AI malware in ransomware attack on critical infrastructure by 2028

Britain's National Cyber Security Centre (NCSC), part of the GCHQ intelligence agency, judged in its annual review that a ransomware attack on UK critical national infrastructure is "highly likely" to involve AI powered malicious software by 2028, per The Guardian. The UK counts energy, transport, defense, food, health and finance among the sectors that infrastructure covers. The review said Russia and China will start using AI routinely in cyberattacks. AI will help attackers scout targets, find weaknesses in IT networks and write malicious code, the agency said. NCSC Chief Executive Richard Horne was set to warn in a speech that security defenses, particularly for critical infrastructure, are inadequate, according to the Financial Times.

Read at The Guardian ↗ • Read at Financial Times ↗ • Read at NCSC ↗

OpenAI to let utilities cut power to future U.S. data centers it develops in grid emergencies

OpenAI launched a Grid Protectors Program under which future U.S. data center campuses it directly develops will agree with utilities on how much electricity they can shed, how quickly and for how long during extreme weather or emergencies, Axios reported. The pilot is OpenAI's planned 3.2 gigawatt campus in Effingham County, Ga., where a 25 year deal approved in August lets Georgia Power cut up to 1,000 megawatts, roughly the demand of 700,000 homes, per The Next Web. OpenAI has committed to cover all grid infrastructure and electric service costs for the Effingham site. "Families and small businesses should not be asked to subsidize a computeComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. campus's electricity needs," the company wrote. Georgia Power projects its agreements with large power users, including OpenAI's, and expected growth will save customers about $950 million a year starting in 2029.

Read at Axios ↗ • Read at The Next Web ↗ • Read at OpenAI ↗

Study finds chatbots gave wrong or outdated voting answers ahead of Election Day

Seven popular chatbots tested by Dewey Square Group for the nonprofit Innovating for the Public Good gave outdated or inaccurate answers to about 14% of questions on voting logistics in October testing, NPR reported, and about 6% of answers could have misled a voter. Three-quarters of the clear errors came from Anthropic's Claude and Microsoft's Copilot. Claude's answers to voting questions were wrong or outdated roughly one time in three, compared with one in four for Copilot and one in 10 for OpenAI's ChatGPT. In a separate NPR test of five chatbots, all declined at first to recommend candidates, but all except Google's Gemini produced lists after two or three prompts.

Read at NPR ↗

New York Times sues Brave, alleging browser maker sells scraped news to AI companies

The New York Times sued Brave Software in federal court in Manhattan, alleging the browser maker copied its articles and distributed them to paying customers including Anthropic, Perplexity and xAI, per TechCrunch. None of the aformentioned customers are listed as defendants. The complaint calls Brave an AI "fence" that sells stolen content to third parties and says Brave used an unnamed crawler, which publishers cannot block, to index more than 30 billion web pages. The New York Daily News, Chicago Tribune and Orlando Sentinel filed a similar suit, saying AI companies pay Brave for access to its index to train models and answer user queries, according to The Verge. The suits also target a Brave feature that lets users get around paywalls, and the Times is seeking an injunction and damages.

Read at TechCrunch ↗ • Read at The Verge ↗

II.China Watch

China's economic planner pledges AI safety rules covering full model life cycle

China plans a broad set of AI safety standards running from development through deployment, including measures meant to guard against loss of control over powerful systems, Bloomberg reported. Wang Ruomeng, a senior official at the National Development and Reform Commission, the state economic planner, said at a Beijing briefing that China will speed up legislation on data, computing power and algorithms and regulate key applications. Wang said safety standards and ethical guidelines will cover the full life cycle of AI, and Beijing will plan a transition period and guidance to ease compliance for AI companies, along with sector-specific regulatory sandboxesSandboxIsolated computing environment where an agent can run code and use tools without touching the machine underneath. Developers and operators would be expected to disclose safety risks, incidents and vulnerabilities.

Read at Bloomberg ↗

Nvidia tells overseas cloud partners to keep Chinese customers a minority

Nvidia has told cloud computing providers that a majority of their services must be sold to non-Chinese companies, the Financial Times reported, citing unidentified people. The requirement, that at least 51% of services go to non-Chinese companies, was relayed verbally by Nvidia’s sales and regional teams and applies only to future contracts. A Singapore-based group said it is difficult to find enough non-Chinese customers outside the U.S. to meet the requirement.

Read at Financial Times ↗ • Read at Bloomberg ↗

OpenAI official urges Washington to guard what it shares in planned AI talks with Beijing

In a Hudson Institute conversation, Caroline Zier, OpenAI's deputy lead for national security, said Washington should be “clear-eyed” about what it shares with China in upcoming bilateral talks on AI, the South China Morning Post reported. "Be very aware of what information you are able to expect in return for what you are sharing, and start to build a foundation," she said. Zier backed holding the dialogue but questioned whether the two governments understand AI risks the same way. She also warned of alleged Chinese distillationDistillationTraining a smaller or cheaper model on the outputs of a larger one so it inherits much of the larger model's ability. It is a standard technique, but it has become a trade dispute because it lets a lab reproduce a rival's capabilities without matching its compute spend. attacks, in which developers train their own models on the outputs of U.S. systems.

Read at SCMP ↗

Huawei AI supercomputer system wins World Internet Conference award as company cites more than 1,000 installations

Huawei's Ascend 950 SuperPoD, which links large numbers of its AI chips to work as one machine, won a Leading Technology Award at the World Internet Conference in Wuzhen, where the company said it has deployed more than 1,000 Ascend SuperPoD systems, per Guangming Online (state media). Buyers include internet, telecom, finance, education and energy companies. Huawei had already said in September that deployments of its Atlas 900 A3 systems topped 1,000. The company said its newest 1,024 chip system can reach 45% utilization of its computing power during training, which it called 1.8 times the industry level, a figure no independent benchmark has confirmed. More than 40 large language models have completed pretraining on Ascend chips, Huawei said.

Read at Pandaily ↗ • Read at Guangming Online ↗ • Read at Huawei ↗

Moore Threads chip cluster in Wuxi runs above 90% utilization after commercial launch, operators say

A computing center in Wuxi built on 2,112 AI chips from Chinese designer Moore Threads entered commercial service in October and runs at 92.9% utilization, according to figures the operators gave Chinese outlet Zhidx. The site took on 55 customers and 33 models during its trial, with training making up more than 62% of the work. In one test running a DeepSeek model, the cluster held 48 concurrent requests within its service target, which the report put at 92% of a setup using Nvidia's H800, though the H800 is a 2023 chip two generations behind Nvidia's current hardware. Self-driving delivery vehicle maker Neolix plans to train its models at the site. Moore Threads is also working with JD Cloud toward a 100,000 chip cluster.

Read at Pandaily ↗

Shenzhen AI chipmaker raises about $150 million, its second round in five months

Moffett AI, a Shenzhen developer of chips that run trained AI models, said it had raised nearly 1 billion yuan ($150 million) in Series C+ funding, TechNode reported. The round follows one of similar size about five months earlier. Moffett's chips use sparse computing, a method that skips unneeded calculations to cut the work of running a model. The company said it has deployed inferenceInferenceRunning a trained AI model to answer a query, as distinct from training it. Inference is where most ongoing compute and energy demand now sits, and it is what chip export rules increasingly target alongside training. clusters, which serve AI models to users, in four regions of China and works with telecom operators. The money will fund upgrades to its current chips and research on next generation processors. Its next accelerator card is due before the end of the year.

Read at TechNode ↗

III.Capability & Research Watch

Google report says attackers hijacking corporate AI access rack up $100,000 a day in cloud charges

Attackers who take over companies' access to large language models are running up fraudulent cloud charges of more than $100,000 in a single day, a report from Google's Mandiant unit and Google Threat Intelligence Group found, per SiliconANGLE. At some organizations the bills reached $1 million. The report, "The Rise of LLMJacking: How Threat Actors Monetize Compromised Cloud and Model Resources," traces most intrusions to stolen or exposed cloud credentials. Other attacks hit exposed servers running Ollama, a free tool for running AI models on a company's own machines.

Read at SiliconANGLE ↗ • Read at Google Cloud ↗

OpenAI more than quadruples typical bug bounty rewards for its AI products

OpenAI raised payouts in its Security Bug Bounty program, with the midpoints of reward ranges for medium, high and critical findings more than quadrupling on average, according to Help Net Security. Critical flaws in ChatGPT, Codex and its API Platform can earn base bounties of up to $30,000, while exceptional critical findings can reach $100,000, or $172,000 with bonuses. The company added a multiplier for high quality reports, a five-tier loyalty program and bonus pools for testing newly launched features. OpenAI will pay the difference on bounties awarded in the 30 days before the change if they would have earned more under the new ranges.

Read at Help Net Security ↗ • Read at OpenAI ↗

IV.Industry & Market Watch

OpenAI names former Army Secretary Driscoll to advise on AI for national security

OpenAI created a national security post and filled it with former Army Secretary Dan Driscoll, CNBC reported. As senior counselor for national security modernization, he will advise company leadership and work with government, industry and academic partners on how advanced AI can strengthen national security and cybersecurity, OpenAI said. Driscoll resigned as Army secretary in August after repeated conflicts with Defense Secretary Pete Hegseth, per Bloomberg Law. Driscoll said in a statement that he is joining to help the U.S. and its allies use the technology responsibly for those who protect them.

Read at Bloomberg Law ↗ • Read at CNBC ↗ • Read at Investing.com ↗

Anthropic amended prospectus shows $17 billion quarter and about $895 billion in commitments

Anthropic’s amended IPO prospectus shows $17 billion in revenue for the quarter ended Sept. 30, more than 11 times the figure a year earlier, SiliconANGLE reported. Quarterly revenue grew 47% from the prior quarter, down from 144% growth three months earlier. The quarter's net loss was $23 billion, per The New York Times. The filing shows about $895 billion in outstanding infrastructure commitments, with leases from Google and Broadcom making up nearly a third and deals with Amazon, Oracle and SpaceX each topping $100 billion. Unexpectedly high demand also forced Anthropic to pay extra for last-minute hardware, resulting in a $5 billion charge.

Read at The New York Times ↗ • Read at SiliconANGLE ↗ • Read at SEC ↗

Marvell agrees to buy Lumentum for $115 billion in stock to expand AI data center networking

Marvell Technology agreed to acquire optical component maker Lumentum Holdings in an all-stock deal valued at $115 billion, per SiliconANGLE. Lumentum shareholders will receive 4.5 Marvell shares for each share they own, about $1,132 a share and a 23.9% premium to the last unaffected price. The deal is a bet on optical networking, which uses light to move data between chips and is increasingly used in AI data centers, The New York Times reported. The companies expect the deal to close in the first half of 2028, according to CNBC.

Read at The New York Times ↗ • Read at SiliconANGLE ↗ • Read at CNBC ↗

Oracle launches $20 billion stock and bond sale to fund AI buildout

Oracle is selling $10 billion in common stock and $10 billion in bonds to fund its AI ambitions, according to separate filings with the Securities and Exchange Commission, The Wall Street Journal reported. Neither offering depends on the other closing, and proceeds may go to capital expenditure, debt repayment, investments or acquisitions, the preliminary prospectus says. Oracle carries about $236 billion in consolidated debt, which ranks ahead of the new common stock. Oracle told investors last Wednesday that new capital markets funding through fiscal 2030 would not exceed $20 billion.

Read at The Wall Street Journal ↗ • Read at Bullished.co ↗

NHS England lists 29 hospital trusts as benefiting from Palantir platform they say they are not using

The 29 hospital trusts that NHS England publicly lists as seeing benefits from Palantir's £330 million ($440 million) federated data platform say the system is not live or that they use none of its applications, The Guardian reported. A senior health official called the gap between the NHS claim and the trusts' accounts "very misleading." The Office for Statistics Regulation, the UK's official statistics watchdog, said it would take up the claim under its casework function. Palantir UK and Europe chief Louis Mosley has said ending the contract would be "putting politics above patients."

Read at The Guardian ↗

V.Global & Geopolitics

High Court judges fault UK prosecutor over made-up precedent that Guardian ties to AI search

A senior lawyer for England and Wales’ Crown Prosecution Service (CPS) cited a nonexistent legal precedent in a High Court case, The Guardian reported, citing court documents. Ravinder Dhillon was defending the CPS’s decision to prosecute a boy accused of raping his younger sister. The Guardian understands the fake citation came from an AI-generated summary in an internet search. Two senior High Court judges called her explanation “unsatisfactory,” though the court said the error did not affect the outcome. It is the second known instance this year of the CPS citing nonexistent authorities. The agency says its policy bars AI from making legal decisions and requires human checks for accuracy.

Read at The Guardian ↗

Brookfield and Drax team up on Yorkshire AI data center that could draw over $27 billion

Brookfield plans an AI data center at Drax's power station in North Yorkshire, CNBC reported. Brookfield said it will initially invest £2.5 billion ($3.3 billion) through its joint venture with Qatar's Qai, and total investment could exceed £20 billion ($27 billion), per ITV News. The first phase targets about 100 megawatts from early 2028, subject to planning and regulatory approvals, with another 500 megawatts added between 2028 and 2031. The project would draw on the station's 4 gigawatts of grid access, and public consultation on the first phase is expected to begin this month.

Read at CNBC ↗ • Read at ITV News ↗ • Read at Yahoo Finance UK ↗