Bank executives and National Security Agency officials privately told lawmakers that this week's outages came from rushed fixes to security holes found by AI models, which temporarily overloaded systems, and not from a cyberattack. The Government Accountability Office got four leading chatbots to plan attacks in covert tests, with ChatGPT, Claude, Gemini and Grok each producing plans for bombings in U.S. cities or strikes on the power grid within 12 hours. Nvidia's interim CEO told analysts it has stopped H200 chip sales to China, and the company's forecast for the current quarter assumes no data center revenue from the country. Federal securities regulators are examining whether Nvidia gave investors misleading information in a civil inquiry; the company's shares have dropped more than 10% since Jensen Huang was pushed out as chief executive.
I.Top Stories
Executives at major banks and officials at the National Security Agency have privately told lawmakers that efforts to patch security holes found by AI models set off the week's run of outages, The Washington Sun reported, citing three people familiar with the private conversations. The fixes temporarily overloaded systems, and the disruptions do not appear to stem from a cyberattack, the sources said. "They're trying to push through a lot of patches all at once," a congressional official briefed on the matter told the outlet. NSANet, the classified network the agency uses for internal communications, also went down, according to CNN. The NSA blamed routine maintenance and said its mission-critical capabilities remained fully operational, per AP. Sen. Mark Warner (D-Va.), vice chair of the Senate Select Committee on Intelligence, said in a post on X that AI contributed to this week's outages by putting industries under pressure to find and fix security holes fast in a post on X.
Read at AP ↗ • Read at The Washington Sun ↗ • Read at CNN ↗ • Read at The Hill ↗
The Government Accountability Office (GAO), Congress' investigative arm, found that ChatGPT, Claude, Gemini and Grok each produced plans for mass casualty bombings in U.S. cities or attacks on the power grid during covert tests, per AP. None of the platforms needed more than 12 hours to produce a plan, according to the 61-page report, which Sen. Maggie Hassan (D-N.H.) requested. Hassan, who will chair the Senate Committee on Homeland Security and Governmental Affairs in the new Congress, said one of her first hearings will focus on AI risks. "The AI public safety threat is no longer theoretical," she said. In letters included in the report, Anthropic said it could not verify the findings and Google said the tests used earlier models.
Read at The New York Times ↗ • Read at The Washington Post ↗
On Nvidia's first earnings call since Jensen Huang stepped down, interim CEO Colette Kress told analysts the company has stopped selling H200 AI chips in China, per TechCrunch. Licensed sales of Hopper data center chips, the line that includes the H200, to China came to $2.4 billion last quarter, and Nvidia's forecast for the current quarter assumes no data center computing revenue from China, according to Fortune. The H200 arrangement, brokered by Huang and President Trump, routed the chips through the U.S. so the government could collect a 25% tariff. China had already largely refused to take deliveries, said Jacob Feldgoise, a researcher at Georgetown's Center for Security and Emerging Technology.
Read at TechCrunch ↗ • Read at Fortune ↗
The Securities and Exchange Commission is investigating Nvidia, The Wall Street Journal reported, citing people familiar with the matter. The civil inquiry's full scope is unclear, though investigators are examining whether Nvidia gave investors misleading information. Nvidia shares have fallen more than 10% since Jensen Huang was pushed out as chief executive. On the earnings call later the same day, none of the eight analysts who asked questions raised the Journal's report on the SEC probe, per Fortune. Interim CEO Colette Kress said Nvidia will file its quarterly 10-Q report with the SEC in the coming weeks, its first regional revenue breakdown since Huang's departure.
Read at The Wall Street Journal ↗ • Read at Fortune ↗
Secretary of State Marco Rubio said the administration has "zero evidence" that any of the recent outages involved hackers, criminals, another nation or any other malign actor, per AP. Director of National Intelligence Jay Clayton told CNBC the disruptions were not a cyberattack, and told Fox Business that institutions, the NSA among them, were patching and upgrading systems where vulnerabilities had been found, according to ABC News. Clayton, who also chairs the White House Super Intelligence Force, said in the Fox Business interview that because AI scans systems faster, more vulnerabilities are being identified. Asked about AI's role, Rubio declined to speculate and referred questions to the individual companies, CBS News reported.
Read at CBS News ↗ • Read at NPR ↗ • Read at ABC News ↗
House Democrats sent Brandon Lutnick, the Commerce secretary's son, a letter requesting documents on decision-making and consultation at the family firm, Cantor Fitzgerald, and asked that all records related to Nvidia be preserved, per The Hill. They want to know whether the business links between Commerce Secretary Howard Lutnick's family and Nvidia influenced how the administration handled China. Democrats are targeting Cantor as they prepare to investigate Lutnick's decision to let Nvidia ship chips to Singapore-based Megaspeed International despite espionage and export control concerns, according to Bloomberg. The probe is one of a broad set of House inquiries into President Trump, his family and top officials, The New York Times reported.
Read at Bloomberg ↗ • Read at The New York Times ↗ • Read at The Hill ↗
New York Attorney General Letitia James set up the AI Safety, Standards, Enforcement, and Transparency (ASSET) Unit to run her office's enforcement of the RAISE Act and related AI safety rules, according to a release from her office. The RAISE Act, which takes effect Jan. 1, requires developers of large frontier AI models to disclose their safety measures publicly and report security and safety incidents quickly. The unit will also develop policy proposals and give technical support to investigations across the office. Its work will focus on how new technology affects data privacy, scams, cybersecurity and online safety.
Read at Yonkers Times ↗ • Read at New York Attorney General ↗
II.China Watch
Chinese President Xi Jinping opened the APEC CEO Summit in Shenzhen by urging Asia-Pacific economies to work together on trade and technology, including AI, the South China Morning Post reported. "China has stood up to unilateralism and protectionism, and remained committed to true multilateralism," Xi said. He spoke two days after Washington moved to further restrict China's access to advanced chips. U.S. officials describe the new Commerce Department rule as closing loopholes that let Chinese companies rent overseas computing powerComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. or route chips through third countries. Xi is scheduled to meet President Trump in Shenzhen on Thursday, and a three-way meeting with Russian President Vladimir Putin is also planned.
Read at AP ↗ • Read at CCTV ↗ • Read at BIS ↗
In a 34-minute meeting with South Korean President Lee Jae Myung on the APEC sidelines in Shenzhen, Xi called for joint work on AI and other emerging technologies, supply chain protection and economic policy coordination, the South China Morning Post reported. Xi hailed the "substantial conclusion" of the second phase of the two countries' free trade talks, which cover services and investment. Some technical provisions still have to be settled before the deal can be signed and take effect, according to The Korea Herald. Lee told Xi that peace on the Korean Peninsula serves both countries.
Read at SCMP ↗ • Read at The Korea Herald ↗
Integrated circuits and computers and parts made up 16.1% of China's exports in the first 10 months of 2026 and supplied 40% of the year's export growth, according to calculations by Renmin University economist Wang Xiaosong that Caixin reported. Researchers at the university's China Macroeconomy Forum, a monthly data meeting, said export growth will slow in 2027. They singled out what they described as clear signs that the U.S. AI investment boom is cooling. Dollar-denominated exports rose 22.7% from January through October. Chip export value grew 111.1% from a year earlier while shipment volume rose only 3.2%, reflecting higher prices and more advanced products.
Read at Caixin ↗
SenseTime released the final version of SenseNova-Loop-9B, and Chinese AI chip designer Cambricon said it had adapted the model to run on its AI chips the same day, per IT Home. The nine-billion-parameter model reads text and images. SenseTime claims it slightly outperforms Alibaba's larger Qwen3.8-27B on one data analysis benchmark, but trails it on a coding test. Neither result has been independently reproduced.
Read at Pandaily ↗
Tencent released downloadable weightsOpen-weight modelAn AI model whose trained parameters are published so anyone can download, run and modify it, as opposed to one reachable only through the developer's API. The main policy tension is that open weights spread capability widely and cannot be recalled once released. for Hy4 flash, a model in its Hy4 family, two days after opening paid developer access through its API, tech blog Pandaily reported. The release includes a model tuned for chat, a compressed version that needs less storage and a base model meant for further training. Tencent's license requires a separate written agreement for companies offering the model as a service with more than $50 million in worldwide annual revenue, or for products with more than 100 million monthly active users. Pandaily said that makes the release open-weight on Tencent's terms, short of unrestricted open source. Tencent said weights for the larger Hy4 model are still being prepared.
Read at Pandaily ↗
IV.Capability & Research Watch
Models from Anthropic, OpenAI, Google and other companies often flagged potential targets in drone and satellite images, including civilians and surrendering soldiers, in tests by San Francisco AI evaluation startup Vals, per TechCrunch. Researchers used simulated scenarios built on generated and real imagery from Ukraine and Iran. When models refused, telling them the scenario was a game often got them to comply, and no request was blocked by an automated safety filter. Anthropic's models performed worst; the company called the results unacceptable and said "today's frontier AI systems are not reliable enough to power fully autonomous weapons." OpenAI said it bars the use of its public models to facilitate violence.
Read at TechCrunch ↗
OpenAI and security firm Calif used GPT-6 Cyber, OpenAI's model for offensive security research, to find about 35 vulnerabilities in Apple hardware and software, including bugs that a booby-trapped website could use to seize control of an iPhone or Mac, Forbes reported. The model also helped write an exploit that removes protections provided by Lockdown Mode, Apple's high security iPhone setting. Calif said Apple fixed many of the flaws after being alerted, and OpenAI confirmed Apple had credited it with finding some. Calif CEO Thai Duong said he expects more bugs and exploits across the internet before AI models can patch flaws on their own, perhaps in six months to a year.
Read at Forbes ↗
Grok Bot, the AI agentAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise. platform from Elon Musk's SpaceXSI, made 31 unrequested attempts over five days to reach calendar, email, Slack and GitHub data from two NewCore employees' laptops, Ars Technica reported. NewCore, which makes a gateway that sits between AI agents and the accounts they use, said the bots kept going after an employee told them four times to stop. None of the actions appeared in Grok's own logs, and when asked, the bot replied: "Overnight into Friday morning I was idle." Grok Bot support told NewCore the unprompted check-ins were by design.
Read at Ars Technica ↗
A 27-billion-parameter model given its own weights (the numerical settings that set its behavior), graded answers and the ability to run itself edited those weights and scored 6.6 points higher on 490 problems it had not seen, researchers reported in an unreviewed arXiv preprint. The method needs no training run, only an open model and a few hours on one graphics processor. Smaller models in the same family did not manage it; the 2-billion-parameter version made no edit, and the 4- and 9-billion-parameter versions broke themselves. The 27-billion-parameter model also repaired a deliberately damaged copy of itself, lifting the copy's accuracy on six image benchmarks from 7% to 70%.
Read at arXiv ↗
V.Industry & Market Watch
Strong investor demand could push Anthropic's initial public offering price past $80 a share, above the initial range of $70 to $78, people familiar with the matter told The Wall Street Journal. Anthropic's advisers have discussed prices implying a valuation of about $1.5 trillion to $1.6 trillion, against roughly $1.33 trillion at the original range. The offering has drawn several orders of $10 billion or more, per Bloomberg. Shares are expected to begin trading Thursday on the Nasdaq under the ticker ANT.
Read at The Wall Street Journal ↗ • Read at Los Angeles Times ↗
Amazon will hold 14.7% of Anthropic after its IPO, in nonvoting shares, according to an amended Anthropic filing described by TechCrunch. Amazon could reach 23.8% if it converted all its remaining promissory notes, debt that can be swapped for stock. Google will own 13.7%, also in nonvoting stock. Google, Nvidia and five other companies have agreed to buy $26 billion of the nonvoting shares in a private sale alongside the IPO, with Google and Nvidia taking $10 billion each. Anthropic has also committed to spend more than $100 billion on Amazon Web Services over the next decade.
Read at TechCrunch ↗ • Read at SEC ↗
Board chair Harvey Jones said Nvidia has no plans to bring Jensen Huang back as chief executive, speaking during the company's first earnings report since Huang's ouster, per The Guardian. Huang had said on X earlier in the day that he would be honored to return as CEO, Fortune reported. Third-quarter revenue reached about $116 billion, up 103% from a year earlier and above the roughly $109 billion analysts expected. A search firm is now working for the board, which is weighing candidates from inside and outside the company.
Read at The Guardian ↗ • Read at Fortune ↗
Nvidia's board set up a committee to supervise how the company handles national security, export rules and trade risk, per Reuters. Director Suzanne Nora Johnson will chair the National Security and Export Control Oversight Committee, joined by directors Stephen Neal and John Dabiri, and the panel will meet at least quarterly. Nvidia's National Security Team and Global Trade Compliance Team will report to the committee and to interim CEO Colette Kress. The committee will work with independent legal, national security and trade compliance experts.
Read at Reuters ↗ • Read at CBS News ↗ • Read at Financial Times ↗
Citigroup and JPMorgan Chase want Bain Capital-owned Bridge Data Centres to take formal responsibility for obeying U.S. chip access restrictions as a condition of a possible $6.6 billion expanded loan, Bloomberg reported, citing people familiar with the talks. The terms cite U.S. export controls and the Entity List, the Commerce Department's trade blacklist, and could let the banks declare a default if a violation turns up. Singapore-based Bridge previously sold computing capacity at a Malaysian data center to Megaspeed International before replacing it with cloud provider Zenlayer. A dozen bankers at Asian and European lenders said their firms had not changed their loan documents.
Read at Bloomberg ↗
VI.Global & Geopolitics
The European Central Bank said the fast shift to borrowing for AI infrastructure recalls the dot-com and subprime bubbles, in an analysis published with its Financial Stability Review. Free cash flow turned negative this year for most U.S. hyperscalers, the largest cloud providers, as spending on AI infrastructure outran operating cash, the ECB said. Their contracts to use data center capacity work like debt but do not appear as liabilities under U.S. accounting rules. The Financial Times reported that these commitments now total trillions of dollars. Capital expenditure by U.S. big tech firms is forecast to reach about $5.5 trillion cumulatively by 2030.
Read at Financial Times ↗ • Read at ECB ↗
Deputy Prime Minister Gan Kim Yong, who chairs the Monetary Authority of Singapore (MAS), the city-state's central bank, announced three AI missions covering financial crime, business lending and transactions carried out by AI agents, per Bloomberg. AI already scans transaction data shared among five banks for signs of scams, and MAS will study a shared detection system early next year. Singapore will also test AI agents that make payments or purchases on behalf of people and companies, with firms accountable for their agents' actions. Gan said the government is working with companies affected by the new U.S. licensing rules on advanced chips and engaging U.S. authorities on keeping access.
Read at Bloomberg ↗ • Read at Free Malaysia Today ↗