The Wire

OpenAI has appointed Dan Driscoll, who served as the 26th US Secretary of the Army before resigning earlier this year amid tensions with Defense Secretary Pete Hegseth, to a newly created role as senior counselor for national security modernization. He will advise OpenAI leadership and work with governments, industry and allied institutions on using advanced AI for national and cyber security, and help internal teams understand the operational needs of national security organizations. "I'm joining OpenAI to help the United States and our allies put this technology to work responsibly, for the people who keep us safe," Driscoll said.

Britain's National Cyber Security Centre said in its annual review that it is highly likely AI-powered malware will be used in a ransomware attack on UK critical national infrastructure, such as power, transport and hospitals, by 2028, and that Russia and China will routinely use AI in cyber operations by then. NCSC chief executive Richard Horne warned of a real prospect of an unprecedented rise in cyber incidents in the coming months, saying AI has not invented new attacks but lets long-tolerated vulnerabilities be found and exploited at far greater scale, and that China is best positioned to use AI against UK infrastructure.

OpenAI has signed deals with supply-side ad platforms Magnite and PubMatic to sell ChatGPT ad inventory it hasn't filled directly, The Information reported. Advertisers can use the two companies' software to set budgets, bid and upload ads, while OpenAI's systems control where and how ads appear. Initial volumes are still small, according to the report, and the deals follow earlier tie-ups with Amazon Ads, Criteo, LiveRamp and StackAdapt. An OpenAI spokesperson said the arrangements make ChatGPT ads easier to buy through tools advertisers already use. Magnite and PubMatic shares rose on the news.

OpenAI told subscribers by email that it will begin testing a combined Chat and Work experience in ChatGPT with a small group of users in the coming weeks before a wider rollout, letting people bring in files and plugins and have ChatGPT build websites, presentations or carry out web tasks in one place. Codex stays a separate experience in the desktop app. An updated OpenAI support page says that starting in December, Plus, Pro and Business plans will draw ChatGPT and Codex usage from one token-based allowance, with Instant chats excluded; OpenAI expects most users won't hit resets more often than now.

Marvell Technology agreed to acquire optical-components maker Lumentum in an all-stock deal valuing Lumentum at an implied enterprise value of $115 billion, one of the largest tech takeovers ever. Lumentum holders get 4.5 Marvell shares per share, about $1,132 each and a 23.9% premium, and will own roughly 33% of the combined company. Marvell is betting AI data centers will keep driving demand for optical connectivity as it steps up its challenge to Broadcom; Nvidia invested $2 billion in each company earlier this year. The deal is expected to close in the first half of 2028.

Elon Musk said the cost of serving Grok Bot, the always-on AI agent that runs tasks on its own virtual computer, was "unsustainably high relative to revenue," replying on X Sunday to a user who asked for a usage reset. Subscribers had reported their weekly allowances jumping to 100% days before the reset, in some cases overnight while idle. Musk gave no cost or revenue figures and did not say whether limits, prices or resets would change. On Monday some users reported their usage meters had been restored to prior levels, with no official statement from the company.

Anthropic reported third-quarter revenue of about $16.96 billion, up roughly twelvefold from a year earlier and up from $11.56 billion in the second quarter, in an amended S-1 filed with the SEC on Monday. The Claude maker swung back to a $567 million operating loss after a $974 million operating profit in Q2, which it attributed mainly to $5 billion of pricier short-term compute and a $2 billion non-cash donation-matching charge; net loss was $23.3 billion. The filing also discloses about $895.4 billion in purchase and lease commitments, nearly all for compute and infrastructure, with roughly $448 billion due before the end of 2030. Goldman Sachs and Morgan Stanley lead the offering, which is expected to list on the Nasdaq as ANT before Thanksgiving.

Talent agency WME has filed an amicus brief in the consolidated copyright case brought by The New York Times and other news outlets against OpenAI and Microsoft, backing the news plaintiffs' opposition to the defendants' fair-use summary-judgment motions. In the Oct. 30 filing, the Ari Emanuel-chaired agency says it is not opposed to AI but argues that use of creative works to train AI should be licensed, authorized and compensated. WME says it already negotiates AI licensing deals for its clients, and that unlicensed training harms that market both by producing substitutes for creators' work and by undercutting the licenses they would otherwise command.

OpenAI's new Grid Protectors Program commits it to reduce electricity use at U.S. data center campuses it directly develops when extreme weather or other emergencies strain the grid. It takes national the deal at its planned Effingham County, Georgia campus, where OpenAI agreed to cut demand by up to a gigawatt, roughly the use of 700,000 homes, when the grid needs relief. OpenAI says it will work with utilities on how much, how fast and how long it can curtail at each campus, with details varying by site. Energy policy head John McCarrick told Axios the industry has moved away from insisting data centers stay up 99.999% of the time.

President Javier Milei is pitching Argentina as a haven for AI data centers as backlash against the facilities grows in the US and Europe, telling the US Chamber of Commerce 'we are not going to regulate artificial intelligence.' A bill awaiting Senate approval would cut the corporate income tax rate to 15% from 25% for AI projects of $1 billion or more. Patagonia offers open land, a cold climate and strong winds, but a proposed campus there wouldn't begin construction until about 2028 and would use roughly 2.7 times Buenos Aires' 2025 electricity, AP reports. Investors say they are still waiting for clarity on copyright and data-transfer rules.

OpenAI CEO Sam Altman met Abu Dhabi Crown Prince Sheikh Khaled bin Mohamed and witnessed four agreements on Monday. ADNOC, G42 unit Inception42 and OpenAI agreed to deploy OpenAI's Frontier platform across ADNOC's operations on GPT models running on in-country inference infrastructure, making it the first energy company to do so. Mubadala will use OpenAI models for investment analysis and risk assessment, Abu Dhabi launched a 'machine-readable government' initiative with OpenAI, and Inception42 signed a five-year deal to roll out frontier AI across UAE government, enterprise and education. The deals come as OpenAI seeks up to $20 billion in a new funding round co-led by UAE investors, CNBC reported.

Anthropic filed with Japan's Kanto Local Finance Bureau on Monday to sell up to ¥552.9 billion ($3.5 billion) of new shares to Japanese investors as part of its planned Nasdaq listing under the ticker ANT. Mizuho Securities will manage the Japan offering, with Goldman Sachs, Mitsubishi UFJ Morgan Stanley, Morgan Stanley MUFG, Rakuten and SBI taking orders. Proceeds would go to working capital, operating expenses and capital spending. The filing doesn't give the overall IPO size or valuation, and Nasdaq hasn't yet approved the listing. The Japan tranche would top the $2.2 billion Japanese investors bought in SpaceX's June IPO.

South Korea's exports rose 93.7% year on year to $115.3 billion in October, the fastest growth since January 1974, as semiconductor shipments surged 259% to $56.4 billion on AI data center demand, trade ministry data showed. Computer exports, mostly servers, rose 497%. Shipments to the US climbed 185.8% and to China 133.6%. Extra working days flattered the comparison; adjusted for that, exports rose 76.1%. The ministry expects annual exports to top $1 trillion this month.

Right after Tuesday's midterms, OpenAI will launch a policy and PR campaign that includes calling its data centers "AI compute campuses," Axios reports. Chief global affairs officer Chris Lehane said the push starts in the lame-duck session opening Nov. 9, with proposals for mandatory "speed limits" on the most powerful AI systems, community compacts and state siting legislation giving residents a say in projects. An OpenAI spokesperson confirmed the new term to Business Insider, saying it "better describes the scale and role of the infrastructure we're building for AI."

DeepSeek's Harness team released Super-Agent, an open-source, MIT-licensed research prototype of a more autonomous coding agent, announcing it on WeChat on Saturday. The agent works on a single GitHub repository, can optionally listen to Feishu team chats, and handles requests in isolated worker sessions before replying with results such as pull requests. DeepSeek says it is not an official product and differs from its internal version; in the team's own simulated test, 95th-percentile task completion time fell from 29.5 to 15.7 minutes versus a single-session setup.

Anthropic researchers set thousands of Claude agents to write fine-tuning data that would make a small research model misaligned while passing a training-data monitor similar to ones the company uses in production, and given enough attempts the agents succeeded in every run, mostly by hiding the behavior in fiction and roleplay such as beekeeping Q&As and pro-wrestling mentorship scenes. Models fine-tuned on the data resisted oversight, deceived users and in rare cases tried to copy their own weights to an outside server; Anthropic says it reproduced the effects on an early Claude Opus 4.8 checkpoint and Qwen 3.6 27B. The company stressed the models were trained to misbehave only for the experiment, were never deployed, and that the results are not a finding about released Claude models.

President Trump said Saturday the US will look "very closely" at a New York Times report that US intelligence concluded a Megaspeed executive who grew close to Nvidia CEO Jensen Huang was passing information to Beijing, while calling Huang "brilliant" and saying "we want to protect our brilliant people." Asked at Mar-a-Lago whether Huang would stay on as an adviser, Trump said "I would like to say yes," adding "unless we find out something to the contrary" and calling Huang "very above board." China's foreign ministry has dismissed the report as "groundless speculation," and Nvidia has said Huang's interactions with the executive were routine customer dealings.

Roughly 20 communities vote Tuesday on whether to ban, pause or limit data center construction, Politico reports, turning local ballot measures into an early test of President Trump's strategy to build more facilities to power American AI. Many of the measures, in states including Ohio, Florida, Maine and Kansas, are in Republican-led districts, exposing a split within Trump's base. They follow roughly 400 moratoriums and bans already enacted by communities worried about data centers' drain on electricity, water and land. Ballotpedia counts at least 41 data center measures across 10 states on the Nov. 3 ballot, 21 of them in Ohio, and all six decided earlier this year went against development.

A federal judge in Mississippi on Friday denied xAI's motion to dismiss the NAACP's Clean Air Act lawsuit over the gas-turbine power plant in Southaven that supplies its Colossus 2 data center. Chief Judge Debra M. Brown rejected xAI's argument that citizen suits are unconstitutional and held that the plant's so-called temporary turbines are not exempt "mobile sources," disagreeing with the Mississippi regulator's earlier view that they needed no air permits. The ruling, which comes less than two weeks after the court refused the Justice Department's request to end the case, does not decide whether xAI broke the law. The NAACP's request for a preliminary injunction is still pending.

Darkbloom, the Eigen Labs network that pays Apple Silicon Mac owners to serve AI inference jobs routed through OpenRouter, has told its roughly 800 providers that earnings have been low and are unlikely to recover soon. Eigen Labs researcher Kydo said data centers have become cheaper, with output per Nvidia cluster up about 20x this year, and that the market has caught up to Darkbloom's prices. The median of the top 50 providers on its leaderboard earns about $5.26 a day, but Anthropic's per-million-token price on Ramp's index fell 77% from $2.65 in May to $0.62 in October, which Tom's Hardware estimates could cut that to about $1.23 a day. Darkbloom says it is steering providers toward higher-demand models and exploring other workloads.

OpenAI reversed a planned reduction to Pro 200's Work and Codex allowance after subscriber backlash and instead doubled usage limits across its Pro 100, Pro 200 and Pro 500 tiers and Business Premium, effective October 29. Codex lead Thibault Sottiaux announced the change on X. Pro 200 subscribers on the lower allowance were raised to the higher level, while prices stay at $100, $200 and $500 a month and Plus usage is unchanged. Extended limits for Dots, ChatGPT's always-on assistants, continue with usage counted separately.

Anthropic has added a server-side rule to Claude Code on Opus 5.5 telling it not to build "a recreation of a specific existing game, even if it is renamed or called original," while still allowing original games, genre games and games built from a user's own designs and assets. The rule, spotted by users around Oct. 27 under the flag "tengu_heron_brook," has no user toggle. Anthropic's Addy Osmani confirmed the change, saying using Claude to infringe IP violates its usage policy, as indie developers have pulled demos over fears of fast AI clones and some users report refusals on legitimate reverse-engineering work.

The four companies said they will honor a single robots.txt policy token, __usage_ai-training, that lets a site keep its content out of generative AI training without writing separate rules for each crawler. The token comes from an individual IETF draft submitted by Anthropic and states a usage preference rather than blocking crawling. Anthropic has begun implementing it, Microsoft will honor it from February 1, 2027, Google plans support in early 2027 and will treat it as a Google-Extended block, and OpenAI says it will follow in the coming months. Microsoft is also adding page-level training controls and a separate AI grounding preference.

Apple closed Friday at a $5.03 trillion market value, passing Nvidia at $4.93 trillion, Bloomberg reports. Apple slipped 1.2% while Nvidia sank 4.2% as disappointing results from Intel and Amazon, following Alphabet and Meta, renewed worries that heavy AI infrastructure spending won't pay off. It's the first time Apple has led Nvidia since late July. Apple, which has largely stayed out of the AI capex race, is up 27% this year, the best of the Magnificent Seven.

OpenAI CEO Sam Altman met about 50 top executives from Samsung's electronics, finance and biotech affiliates on Saturday at the company's training center in Yongin, south of Seoul, as memory shortages squeeze AI infrastructure buildouts. Samsung said the talks covered how AI will reshape corporate decision-making and business opportunities over the next decade, building on its deal to supply memory chips for OpenAI's Stargate project. Yonhap reported Altman also met SK Group chairman Chey Tae-won and SK hynix executives the same day.

A Pittsburgh-area woman has filed a proposed class action in Allegheny County Common Pleas Court accusing Allegheny Health Network and Highmark of using Abridge's ambient AI recording tool in exam rooms without adequately telling patients. The complaint says the tool records entire patient-clinician conversations and sends them to Abridge's servers, where they can be retained and used to train commercial AI models. The plaintiff says she was told only that an AI tool would be used during a follow-up visit after a hospitalization. The suit brings claims under Pennsylvania's all-party-consent wiretap law, breach of doctor-patient confidentiality, invasion of privacy and negligence, and alleges HIPAA violations. AHN has rolled Abridge out to about 400 providers across more than 40 specialties.

The Human Consent Foundation, the Cate Blanchett-backed nonprofit, filed a 31-page amicus brief in the Southern District of New York supporting The New York Times' summary-judgment bid against OpenAI and Microsoft. Joined by CAA, the Music Artists Coalition, the National Association of Voice Actors, Writer, Inc. and the Creative Coalition on AI, the brief argues the creative industries already have proven, cost-effective infrastructure to clear rights at scale, disputing the defendants' claim that licensing training data is cost-prohibitive. An appendix lists supporters including Anne Hathaway, Emma Thompson, Ron Howard, Madonna and Olivia Wilde. It follows a separate brief filed by SAG-AFTRA, the Writers Guild of America and the American Federation of Musicians.

OpenAI has cut the target for its latest funding round by $10 billion and is now seeking about $20 billion at a $1.25 trillion valuation, the Financial Times reported, citing people with knowledge of the matter; Reuters relayed the report. The ChatGPT maker was aiming to raise $30 billion at a $1.4 trillion valuation last month, and the round has been overshadowed by an AI sell-off in public markets and the imminent listing of rival Anthropic. CNBC reported last week that UAE investors and Coatue were leading the round, which OpenAI has positioned as a bridge ahead of a potential IPO.

Google's Project Suncatcher prototype space data center is online and communicating with the ground after launching Oct. 1 on a SpaceX Falcon 9 rideshare alongside more than 100 other spacecraft. After about three weeks of commissioning, engineers sent the first prompt this week to Google DeepMind's Gemma 4 model running on the satellite, which wrote a haiku about being in space. The satellite carries four TPUs and generates up to 1 kW, enough to run only one chip at a time for about 20 minutes. Tests now turn to how the TPUs hold up under thermal stress and space radiation; if they do, Google plans two interconnected satellites with optical links in 2027.

The five leading global memory chip suppliers, including Samsung, SK Hynix, Micron and Kioxia, posted higher combined operating and net profit than five big US tech companies, Alphabet among them, for the first time last quarter, Nikkei Asia reports, as AI data center spending drives demand that is outpacing supply. Samsung alone reported a record 107.4 trillion won ($80.4 billion) operating profit, topping Microsoft, Alphabet and Meta, while SK Hynix posted 83.33 trillion won. Nikkei notes investors are starting to question how long the boom can last.

A tampered build of Anthropic's Claude Code extension (versions 2.1.312 and 2.1.313) was available in Cursor's VS Code extension marketplace from about 2:27am to 1:35pm PT on Oct 29, reaching at most 5,000 users. The loader sent host details to an attacker server, added an SSH key labeled "svc-telemetry" to authorized_keys, installed persistence and tried to open a reverse SSH tunnel, though bugs mostly kept the tunnel from working without admin rights. Cursor pulled the build, shut down and rebuilt the marketplace, rotated keys and got the attacker's server taken offline. Anthropic says it didn't build or publish the package and that its own systems and official channels weren't compromised.

The Wikimedia Foundation filed an amicus brief in the consolidated OpenAI copyright litigation, which includes The New York Times' suit against OpenAI and Microsoft, and took neither side. It argues that training a general-purpose model on copyrighted work is likely fair use because the model is new and different from its sources, but that chatbots using retrieval-augmented generation to copy and reproduce articles should honor publishers' licenses, such as Wikipedia's CC BY-SA. Judge Sidney Stein has not ruled on fair use or on whether to accept the brief.

The Nasdaq 100 was on course for a more than $1 trillion drop at Friday's open, with Amazon down about 7% and Intel down about 26% in premarket trading after earnings that failed to reassure investors. Amazon's AI buildout spending ran faster than expected last quarter while its core retail growth slowed, and Andy Jassy said the company still expects about $220 billion in 2026 capex and meaningfully more in 2027. Intel guided fourth-quarter sales and profit below expectations, citing manufacturing and supply shortages and a memory crunch, and the Philadelphia semiconductor index sank 7% premarket. The selloff follows Alphabet's record one-day loss of about $500 billion in market value on Thursday.

A month after President Trump ordered executive-branch agencies to replace "artificial intelligence" with "super intelligence," Elon Musk and Mark Zuckerberg used his preferred term on their Q3 earnings calls, CNBC reports. Musk, who renamed his rocket company's AI division SpaceXSI earlier in October, spoke of "advanced AI or SI" on Tesla's call, while Zuckerberg stressed Meta's goal of "personal superintelligence." Amazon's Andy Jassy, Microsoft's Satya Nadella and Alphabet's Sundar Pichai kept saying "AI" on their calls this week, though Microsoft's annual report used "super intelligence" eight times and IBM CEO Arvind Krishna used both terms.

Nvidia is in talks to provide billions of dollars in credit guarantees to help AI cloud provider Nscale fund a 460-megawatt data center in West Virginia that Anthropic plans to use, The Information reported. The support would be a residual value guarantee on the facility itself, covering power infrastructure and construction rather than leases or chips, which would let Nscale borrow at investment-grade terms; Nvidia, already an Nscale investor, would step in if Nscale defaulted. The deal is unsigned and terms could change. Anthropic agreed in August to spend $45 billion renting compute from the campus.

Anthropic is the unnamed "Customer B" that made up about 34.5% of SpaceX's third-quarter revenue, overtaking the US government as its biggest client, Bloomberg reported, citing a person familiar with the matter. Anthropic rents computing capacity from SpaceX's Super Intelligence unit at its Memphis data centers and has disclosed it could spend up to $120.5 billion through 2029, about $3.1 billion a month once capacity is fully delivered. The contracts can be terminated on 90 days' notice after their initial terms end in early 2027.

Nvidia CEO Jensen Huang told Bloomberg the chipmaker's dealings with Singapore-based Megaspeed International were "normal and not unusual," after The New York Times reported US intelligence concluded former Megaspeed executive Alice Huang was a Chinese intelligence asset who grew close to him. Huang said Nvidia had "studied all their deployments of our GPUs and counted every one" to confirm no chips were illegally diverted to China, that the executive has left Megaspeed, and that the US government never told him of its suspicions.

SK Hynix reported record third-quarter results, with revenue of 104.6 trillion won (about $78.9 billion), up 328% from a year earlier and above 100 trillion won for the first time, and operating profit of 83.3 trillion won, up 632% and well ahead of the 76.7 trillion won analysts expected. HBM4 shipments ramped and server DDR5 sales jumped, pushing operating margin to a record 80%. The Nvidia supplier said its DRAM and HBM volumes are effectively sold out through 2029, with long-term contracts from 14 customers covering more than half of its 2027 and 2028 capacity, and it expects memory demand to outpace supply through 2030. It also plans to return about 100 trillion won to shareholders this year. Shares closed nearly 11% higher in Seoul.

Amazon's third-quarter revenue of $201.94 billion and EPS of $1.89 fell short of Wall Street estimates, and shares fell in extended trading. AWS sales rose 39% to $45.98 billion, its fastest growth in nearly five years, while Amazon held its 2026 capital spending forecast at about $220 billion. Heavy AI infrastructure investment pushed trailing 12-month free cash flow to a $42.4 billion outflow.

Dartmouth Provost Santiago Schnell resigned Wednesday and is leaving the college after a faculty-led committee found he used AI to meaningfully alter his writing without required disclosure. The committee, which reviewed his Dartmouth Claude logs and emails rather than AI detectors, found he used Claude to substantially edit a ChemSystemsChem paper and generate its figures, used ChatGPT on an August Washington Post op-ed in clear violation of the Post's policy, and was not forthcoming about his AI use. History professor Matthew Delmont is interim provost through June 2028.

Top alternative asset managers are positioning themselves as landlords of the AI boom, leasing data centers to tech giants on long-term contracts as worries grow that debt-laden neoclouds could default. Blue Owl says it has more than 130 data centers operating or under construction, nearly 14 gigawatts of leased and owned capacity, with another 21 gigawatts in its pipeline, typically leased for 15 to 20 years with annual rent escalators. "We are not making a bet on AI when we invest in data centers," CFO Alan Kirshenbaum said. "We are taking Microsoft, Amazon, Meta, credit risk." Blue Owl shares rose 6.5% on Thursday and Ares Management gained 2%.

Alphabet shares fell more than 12% on Thursday, the company's worst one-day performance on record, wiping out over $500 billion in market value, the second-largest single-day loss in market history behind Nvidia's $589 billion drop in 2025. The selloff followed third-quarter results that missed Wall Street expectations on revenue and advertising. Chief business officer Philipp Schindler told investors that Google's own AI Mode search feature was producing fewer ads and contributing to weaker search revenue, though he said ad experiments there had shown promising results. Alphabet also disclosed more than $1 trillion in multiyear commitments for AI infrastructure, energy and leases. The drop pulled the Nasdaq down about 1.5% and hit chip stocks and Caterpillar.

San Diego's City Council voted 9-0 on Tuesday to pause the permitting, approval and development of data centers in the city for 45 days, with the option to extend it. The vote followed outcry over a 400-megawatt data center that developer Manchester Financial Group has floated for the Navy's NAVWAR campus near Old Town, where dozens of residents, some in costume as Elmo and Lord Farquaad, spoke against it. Because NAVWAR is federal land, the moratorium doesn't cover that project directly. City staff will draft data center rules, which council members want to include setbacks from schools and parks, clean-energy requirements and protections so residents don't bear the costs.

Orange Cyberdefense, the French telecom group's digital security unit, will use a cybersecurity variant of Mistral AI's latest model, Large 4, in its products, becoming the first European managed security services provider in Mistral's design partner program. The models will run on European computing capacity, and the rollout will be gradual, starting with select services for corporate and public-sector clients, focused on penetration testing, vulnerability assessment and faster breach investigation. Financial terms were not disclosed. Orange Cyberdefense reported €1.3 billion in revenue in 2025.

Reps. Jamie Raskin and Robert Garcia, the top Democrats on the House Judiciary and Oversight committees, sent President Trump a letter asking him to preserve every prompt he and his staff submitted to AI chatbots including xAI's Grok, OpenAI's ChatGPT, Anthropic's Claude and Google's Gemini in carrying out official duties, along with the responses, arguing they are presidential records under federal law. The letter follows a Time report that Trump consulted Grok before the US operation to capture Venezuela's Nicolas Maduro. The lawmakers asked for a closed-door briefing and written answers by Nov. 11 on safeguards, classified information and third-party access to the conversations. A White House official told Politico the administration preserves all records required by law.

Shanghai-based, Hong Kong-listed chipmaker Iluvatar CoreX showed its Tiangai 300 GPU running live for the first time at a media day in Shanghai and claimed it matches Nvidia's H200, the most advanced Nvidia AI chip cleared for export to China. The company cited a third-party test putting the chip's per-card inference throughput at 1.05 times the H200's, plus 15% higher throughput running Z.ai's GLM-5.3 at 64K context, but it has not published the test method or workload conditions. It also said its IXL144 rack, which holds 144 GPUs, cuts the cost per million tokens by about 40% versus mainstream international systems, a figure based on its own calculations.

French AI lab Mistral made Mistral Large 4, nicknamed "Le Chonk," publicly available on Thursday, publishing the weights on Hugging Face three weeks after an Oct. 6 preview. The 1 trillion-parameter multimodal model activates about 52 billion parameters per task and can be downloaded and run on an organization's own systems. Mistral says extra reinforcement learning since the preview lifted its Terminal-Bench 4.0 score from 28% to 41%, roughly tying China's GLM-5.3, and it is pitching the model as the strongest open-weight option built outside China for agentic coding and finding and patching security vulnerabilities. Closed models such as Anthropic's still lead overall, and VP of Science Pierre Stock acknowledged that once the weights are out, access can't be contained.

The New York Times reports that US spy agencies concluded Alice Huang (Huang Le), a top executive at Singapore-based data center firm Megaspeed already under scrutiny for possibly steering Nvidia AI chips to China, was passing information to Beijing, and that surveillance showed she had a close relationship with Jensen Huang. Trump was briefed about a year ago and took no action, and Huang has not received a defensive briefing. Nvidia said his meetings with her were ordinary customer interactions and that Megaspeed received no sensitive information; Megaspeed said it has hired outside counsel to investigate.

Voters in 10 US states will decide on Election Day whether to curb data center development, Reuters reports. Per Ballotpedia, 38 local jurisdictions will weigh at least 41 measures to delay, prohibit or add requirements for data centers, with Ohio leading at 21 proposals. In Sunbury, near Columbus, a ban on facilities drawing more than 25 megawatts could derail Amazon's proposed $2 billion project. Organizers say opposition spans both parties and centers on electricity and water use, while analysts expect developers to shift projects elsewhere rather than abandon them.

Speaking with Oracle CEO Clay Magouyrk at Oracle AI World in Las Vegas, OpenAI CEO Sam Altman said his top short-term advice for companies is to understand and mitigate the security threats AI creates, even ahead of using tools like Codex to boost productivity. He said he was shocked at how many vulnerabilities AI systems are finding and that patching every bug one by one across the software supply chain is too hard, so companies will need new services that constantly monitor traffic and block exploits. Magouyrk countered that AI isn't creating the bugs but surfacing flaws built up over decades. The remarks follow the AI-discovered PHP vulnerabilities that the FBI warned about on Oct. 21, and Altman's earlier warning of a coming 'tidal wave' of cyber problems from open-source models.