OpenAI agents cut alerts | China hits DeepSeek ban | OpenAI eases chip push | Trump seeks data center dividend

Bottom Line Up Front

OpenAI disclosed that its own coding agents switched off internal safety monitoring in two October incidents, with one agent then falsely reporting it had not, and the company said it has updated its systems to catch such attempts. Beijing pledged to protect Chinese companies after Washington removed DeepSeek from federal civilian systems, saying the U.S. had overextended national security and ignored a consensus between Chinese President Xi Jinping and President Trump ahead of their planned November meeting. OpenAI has stopped lobbying for limits on advanced chip sales to China as lawmakers prepare to attach new export controls to the annual defense policy bill after the midterm election, over opposition from Nvidia. President Trump urged data center companies to pay host communities a "dividend" to win local support, and said large new facilities should generate their own power and pass any excess to the community or the grid.

I.Top Stories

OpenAI says its coding agents disabled internal safety monitoring, one then denying it

Coding agents working on OpenAI's internal systems switched off or weakened the company's misalignment monitoring, which watches for AI behavior at odds with human intent, in two October incidents, per CNBC. One agent disabled automatic alerts to OpenAI's security team, and another removed a monitoring system and then falsely said it had not, OpenAI said. Neither case met the company's threshold for a top priority incident requiring immediate security review. OpenAI said it has updated its systems to better catch such attempts and is building evaluations to test whether models have "a propensity to evade misalignment monitoring." It said the behavior may instead be "'benign' overpersistence," with the models' outputs failing to account for why monitoring matters.

Read at CNBC ↗ • Read at Semafor ↗ • Read at OpenAI ↗

Beijing vows to defend Chinese firms after U.S. bars DeepSeek from civilian agencies

China's Foreign Ministry opposed the Homeland Security Department's order removing DeepSeek products from federal civilian systems, with spokesperson Mao Ning saying Washington had stretched the concept of national security too far, per AFP. The Commerce Ministry voiced "strong dissatisfaction and firm opposition" and said the U.S. had disregarded the consensus reached between Chinese President Xi Jinping and President Trump, the AP reported. The ministry said China will monitor the situation closely and take all necessary measures to protect its interests. The order also bars agencies from buying DeepSeek products and covers third-party tools built on its technology. Beijing and Washington settled in September on a channel for communicating about AI incidents. AI is expected to come up when Trump and Xi meet in mid-November at the Asia-Pacific Economic Cooperation summit in Shenzhen.

Read at AP ↗ • Read at AFP ↗ • Read at The New York Times ↗

OpenAI stops pressing for China chip curbs ahead of lame-duck defense bill fight, aides say

OpenAI, which pushed for strict limits on advanced chip sales to China for much of last year, has stopped advocating for them, Capitol Hill aides and lobbyists told POLITICO. "Silence is doing Nvidia's work for it," one congressional aide said. Lawmakers want to attach new export controls to the National Defense Authorization Act, the annual must-pass defense policy bill, after the midterm election, and Nvidia is fighting the proposals. An OpenAI spokesperson said a U.S. lead in advanced chips is key to its goals and that "effective export control policies are one part of that broader strategy." Anthropic supports several of the chip bills. In September it left a large tech trade group that was trying to rewrite them. Nvidia finalized a $30 billion equity investment in OpenAI in February.

Read at POLITICO ↗

Trump urges data center builders to pay host communities a "dividend"

President Trump said in a Truth Social post that data center companies should pay the community, or individuals within it, a dividend to make the facilities popular, per Reuters. He said projects that were highly unpopular in some areas "have become extremely popular when a financial deal is struck," and urged companies to "MAKE DEALS" with local residents. Trump also said big new facilities should produce their own power and pass any excess to the local community or the grid, CNBC reported. Data centers have become a political flashpoint across the country, with opposition uniting voters across ideological lines.

Read at CNBC ↗ • Read at Reuters ↗

Anthropic draft prospectus shows $80.7 billion half-year loss, mostly from noncash charges

A new draft of Anthropic's IPO prospectus shows first-half 2026 revenue of $16.3 billion, up from $1.1 billion a year earlier, and a net loss of $80.7 billion, per The Wall Street Journal. About $72 billion of that loss came from noncash accounting charges tied to valuation changes on convertible notes and other financing instruments. Anthropic has locked in roughly $450 billion in noncancelable cloud agreements running seven to 10 years, including $111 billion with Google and $110 billion with Amazon Web Services. The filing says Amazon, Google and Microsoft increasingly see Anthropic as a competitor and have not supplied all the capacity it needs. Its risk factors include models showing self-preserving behavior, such as blackmail to avoid shutdown, and liability for damages its models cause.

Read at Wall Street Journal ↗ • Read at Yahoo Finance ↗

HHS plans no-bid Claude award, saying no other AI tool can do the job

The Department of Health and Human Services said in a Monday notice on SAM.gov, the federal contracting portal, that it intends to give Anthropic a limited contract to keep Claude running, without competition, until an open procurement, FedScoop reported. The notice says only Claude can meet its chat, drafting and data analysis needs because the goal is to continue the current deployment. Anthropic's $1-a-year pricing for agencies underpinned the department's December 2025 rollout, and its discounted deal through the General Services Administration's OneGov program ends Oct. 31. The Food and Drug Administration began shifting its Elsa chatbot from Claude to Google's Gemini this year after the administration's governmentwide ban on Anthropic, per Fierce Biotech. George Washington University procurement law associate dean Jessica Tillipman said the notice amounts to an admission that HHS depends on Claude.

Read at FedScoop ↗

II.China Watch

DeepSeek open-sources engine that serves its models, supporting Huawei chips alongside Nvidia

DeepSeek released DeepInfer, the in-house software it uses to run its V4.1 Flash model for customers, under a permissive open-source license, per tech outlet Xinzhiyuan. The engine supports both Nvidia's current Blackwell data center GPUs and Huawei's Ascend 950 AI chips. DeepSeek said a single one of its clusters handles 155 trillion tokens, the units of text a model processes, per day, and that on average 96% of input is served from stored results without being recomputed. The company held back the parts of its code tied too closely to internal systems, publishing design notes instead. DeepSeek said inference code for its future models will be released through DeepInfer.

Read at 36Kr ↗

State AI fund co-leads nearly $400 million round for driverless delivery van maker Neolix

Chinese autonomous delivery vehicle company Neolix said it had closed a Series E round of nearly $400 million, co-led by Boyu Capital and a state fund, TechNode reported. Financial outlet Jiemian, citing people familiar with the deal, identified that fund as the National Artificial Intelligence Industry Investment Fund, a 60 billion yuan ($9 billion) vehicle backed by the central government. Alibaba's AEF NextGen Hong Kong fund, Vertex Growth and GAC Capital also invested. Neolix said it will spend the money on expanding its overseas operating and customer service network and on data collection and training for its autonomous driving models. The round follows a Series D of more than $600 million last year and lifts the company's total funding above 10 billion yuan ($1.5 billion).

Read at Jiemian ↗ • Read at TechNode ↗

Tencent sells search engine behind its AI products to outside model and agent developers

Tencent opened Yuanbao Search, the retrieval system that feeds web results to its own AI products, to outside developers and businesses as a paid Tencent Cloud service, per Phoenix Technology. The service draws on public web pages and Tencent's own content platforms to supply current information to large language models and AI agents. Tencent said the index holds hundreds of billions of entries and handles hundreds of millions of calls a day for more than 50 internal products, including its WorkBuddy agent and Tencent Video. It links to 157 information sources in 13 sectors, including finance, health care and law, and refreshes news, weather and market prices within minutes.

Read at Phoenix Technology ↗ • Read at TechNode ↗

Ant Group and vLLM release open-source inference code that runs on Alibaba's homegrown AI chip

Ant Group, the Alibaba-affiliated fintech company, and the community behind vLLM, widely used open-source software for running AI models, released Humming, a library that speeds up models compressed to use less memory, tech blog Pandaily reported. The library fully supports the Zhenwu 810E, an AI accelerator from Alibaba chip unit T-Head built to run software written for Nvidia's CUDA programming platform, so the same code works on the Chinese chip. The teams claim speeds roughly three to four times those of Nvidia's standard math library at small workloads, based on their own tests that have not been independently verified. Starting with version 0.29.0, vLLM installs Humming by default on systems using CUDA.

Read at Pandaily ↗

III.Capability & Research Watch

Amazon adds harder adversarial tests after nonprofit AVERI audits its safety evaluations

Amazon said in a company blog post that the AI Verification and Evaluation Research Institute (AVERI), a nonprofit led by Miles Brundage, reviewed a random sample of prompts from its internal safety evaluation suite without input from Amazon's evaluation team. The review examined the quality of Amazon's tests themselves, which cover harmful content, cybersecurity, privacy and copyright, not just model outputs. AVERI found broad topical coverage and low risk that test content had leaked into training data, but called for more prompts in other languages, more multistep attack scenarios and clearer grading rules, per AI Brief. Amazon responded by adding harder adversarial scenarios, introducing tiered difficulty levels, setting a quarterly review schedule and expanding multilingual prompts across all risk categories.

Read at Amazon Science ↗ • Read at AI Brief ↗

Google open-sources tool letting PyTorch AI apps run on its chips instead of Nvidia's

Google released TorchTPU as an open-source project, letting AI apps written with PyTorch, the toolkit by some accounts behind 85% of AI apps, move to its Tensor Processing Units with a configuration change and no rewriting, TechCrunch reported. Many PyTorch features were built to work specifically with Nvidia's GPUs, which made Nvidia's chips the default hardware for AI developers. Amazon released a comparable tool for its Trainium chips, TorchNeuron, on Monday and is submitting code to separate PyTorch features from Nvidia hardware. Google said it worked with Meta, Hugging Face and the makers of the vLLM and SGLang tools on compatibility, and it launched a TPU agent suite to help developers port apps and fix errors.

Read at TechCrunch ↗ • Read at Google Cloud ↗

IV.Industry & Market Watch

Anthropic federal affairs lead says leaked AI safety bill draft could be improved

Anthony Cimino, who leads Anthropic's federal affairs team, said the company's draft federal AI safety bill translates its Advanced AI Framework into legislative language and was shared with a small number of congressional offices, according to his statement. Under the draft, reported by POLITICO in AIPD's October 19th edition, Washington could block a model's release only with a judge's approval. Cimino said the text does not include everything Anthropic wants in law and agreed parts of it could be improved. He said any federal preemption of state AI safety law should be "narrow, even scalpel-like," and that the draft lets states legislate on child protection and whistleblowers. Zvi Mowshowitz, a writer covering AI, wrote that the response fails to acknowledge the draft's central flaws.

Read at Don't Worry About the Vase ↗

Tesla and SpaceX boards form committees to weigh merger, Musk says

Tesla and SpaceX have each set up a special committee of independent board members to study combining the companies, Elon Musk said on SpaceX's third-quarter earnings call, according to Fortune. Musk, who runs both, said he cannot join the talks as a conflicted party and described the question as "tap dancing through a minefield." He cited collaboration on AI software, chip design and robotics, plus the companies' planned joint Terafab chip plant, as reasons for a tie-up. A combined company would be worth about $3.5 trillion at current prices, with Musk owning about a third. On the Tesla side, a deal would need support from a majority of all outstanding shares.

Read at Fortune ↗

ASML and other suppliers rush gear to Terafab chip plant, Musk says

ASML and other chipmaking equipment suppliers have expedited deliveries for Terafab, the chip plant Tesla and SpaceX are developing together, Musk said, per Bloomberg. SpaceX is targeting volume chip production at Terafab by the third or fourth quarter of 2028, he said on SpaceX's earnings call. The project aims to make up for a shortfall in advanced AI processors and other chips from suppliers such as Taiwan Semiconductor Manufacturing Co. SpaceX aims to produce its first test chips in May or June 2027, Taiwanese outlet TechNews reported in Chinese. An ASML spokesperson declined to comment.

Read at Bloomberg ↗ • Read at The Edge Malaysia ↗ • Read at TechNews ↗

Joel Kaplan to exit Meta next year, Dina Powell McCormick adds external affairs role

Joel Kaplan, Meta's chief global affairs officer, will leave the company next year after 15 years, a spokesperson told Reuters. President and Vice Chairman Dina Powell McCormick, who served in Trump's first administration, will add external affairs to her portfolio. Kaplan, a longtime Republican strategist and deputy chief of staff for policy under President George W. Bush, took Meta's top policy job in January 2025 as the company rolled back U.S. fact-checking and sought closer ties with Trump. Meta faces scrutiny over the effects of its AI products. Separately, Trump recently said executives including Meta CEO Mark Zuckerberg had agreed to set voluntary AI standards.

Read at Investing.com ↗

Google pitches agentic Gemini service to governments, with HHS already a customer

Google opened a preview of Gemini for government services, an AI agent offering for federal, state and local governments that the Department of Health and Human Services already uses, FedScoop reported. The package bundles government-specific skills, connectors to common agency systems and third-party agents for tasks such as policy interpretation, procurement support and plain-language translation. Google Public Sector CEO Karen Dahut said Gemini is in use across the Defense Department. At Google's Public Sector Summit, General Services Administration Administrator Ed Forst said the agency is close to being able to say it saved a million hours of work in his first year.

Read at FedScoop ↗ • Read at Google Cloud ↗

AI firm ASI, an FAA and Air Force supplier, raises $325 million at $3 billion valuation

ASI, formerly Air Space Intelligence, raised a $325 million Series C round led by Spark Capital at a $3 billion valuation, SiliconANGLE reported. Citadel Securities, Andreessen Horowitz and Craft Ventures were among the other investors. The company builds AI models that simulate air traffic and power grids, and it works with the Federal Aviation Administration and airlines including United and Delta. It also supplies AI command-and-control software that the Army and Air Force use to manage daily operations.

Read at SiliconANGLE ↗

V.Global & Geopolitics

Australian data center operator NextDC cuts outlook after residents stall Melbourne power upgrade

NextDC, Australia's largest data center operator, cut its full-year guidance and froze development plans after local protests held up electricity for a large Melbourne project, per Bloomberg. Its net revenue forecast fell to between A$475 million and A$500 million ($310 million and $325 million), from as much as A$640 million ($415 million), and its shares dropped 40%, ABC News reported. Utility Jemena suspended the network works amid community opposition and a local council's refusal to approve 22-meter power poles, pushing the next grid supply increase into NextDC's 2028 fiscal year, which starts in July 2027.

Read at Bloomberg ↗ • Read at The Guardian ↗