Show only each day's banner story

Microsoft restores Outlook, Teams and Copilot after global outage

The outage hit Copilot and Microsoft 365 work services at the same time. The company linked it to network infrastructure.

On November 14 Microsoft said it had fully restored Outlook, Teams, Copilot and other Microsoft 365 services and apologized to users. According to Downdetector, complaints came from more than 70 countries and territories. By the company's account, the outage lasted more than 12 hours and surfaced after a change to network settings affecting a segment where capacity had previously been expanded. Microsoft ruled out any link to a cyberattack or SI agents and promised to publish a report on the investigation's findings.

Ahead of the listing, Amodei singled out one of Anthropic's key risks: securing enough computing capacity to sustain growth.

Dario Amodei calls compute access Anthropic's main dilemma ahead of IPO

Ahead of the listing, Amodei singled out one of Anthropic's key risks: securing enough computing capacity to sustain growth.

Speaking with JPMorgan Chase CEO Jamie Dimon, Dario Amodei called access to computing capacity Anthropic's main business dilemma and said he sees no big problem in choosing the month to go public. According to The Decoder, Amodei also noted that the coming year remains highly uncertain. Per the latest filing, Anthropic's total obligations amount to almost $901 billion, of which about $71 billion is due by September 30, 2027. The offering price is due to be set on November 18, with trading on Nasdaq expected to begin on November 19.

The most valuable company of the SI era is moving into a succession mode: who leads Nvidia will determine the pace of accelerator supplies for the entire industry — and the market reacts to uncertainty more strongly than to chip news.

Nvidia board seeks Huang's successor: FT learns of approach to Arm chief Rene Haas, company denies contacting candidates

The most valuable company of the SI era is moving into a succession mode: who leads Nvidia will determine the pace of accelerator supplies for the entire industry — and the market reacts to uncertainty more strongly than to chip news.

According to the FT, citing sources, Nvidia's board, in its search for a successor to Jensen Huang, approached Arm CEO Rene Haas, who worked at Nvidia for about seven years and was a vice president and general manager of its compute products division before moving to Arm in 2013; Haas is only one of several candidates, and there is no guarantee of an offer or acceptance. Nvidia's board told Reuters it is still selecting a search firm and has not formally approached anyone, calling discussion of candidates premature; the acting CEO is CFO Colette Kress. Arm shares fell about 5%, Nvidia about 2%. According to Chinese business newswires, Nvidia confirmed it has not contacted candidates and has not chosen a headhunter.

The supply chain of SI tools turned out to be the weak link: a key at an inference provider became the right to publish extensions on Anthropic's behalf, and marketplace security for plugins will now have to be built to the standards of software repositories.

Cursor confirms stolen Fireworks AI key let attackers publish malicious Claude Code extension

The supply chain of SI tools turned out to be the weak link: a key at an inference provider became the right to publish extensions on Anthropic's behalf, and marketplace security for plugins will now have to be built to the standards of software repositories.

Cursor has published an incident report: attackers obtained an access key held by inference provider Fireworks AI, used it to copy data from a read-only bucket holding historical traces of SI coding agents (mostly from 2025), some of which contained unredacted API credentials, and used those to access Cursor's internal SaaS systems. The company estimates traces were exposed for fewer than 0.05% of users and API keys were compromised for fewer than 0.01% of customers. Having obtained credentials for publishing to the marketplace, on October 29 the attackers released malicious versions of Anthropic's Claude Code extension, 2.1.312 and 2.1.313, which were available for about 11 hours; up to 5,000 developers may have received builds that install persistent SSH backdoors. Cursor cut off all inference traffic from Fireworks, rotated keys and rebuilt the marketplace; according to BleepingComputer, the legitimate Anthropic and Open VSX extensions were not affected, and the incident is attributed to the "Black October" wave of attacks on SI platforms.

Hacking "by SI agents" has now earned its own FBI warning: detecting and patching vulnerabilities must now be measured in hours, not weeks.

FBI warns of a wave of breaches using the open-source agentic framework ARTEX

Hacking "by SI agents" has now earned its own FBI warning: detecting and patching vulnerabilities must now be measured in hours, not weeks.

The FBI issued alert FLASH-20261113-001 about a surge in automated attacks using the open-source ARTEX framework — a Go-based tool built from interacting SI agents (a "planner" and "executors") with a shared graph knowledge base. In the cases documented by the FBI, attackers went from reconnaissance to privileged access in under three hours; Unit 42 noted stolen credentials being used within seconds. The Amazon campaigns targeted theft of SI model keys to pay for inference; DeepSeek Flash is the most common, and there are more than 1,600 ARTEX consoles exposed on the open internet (Censys). The tool appeared on GitHub in the summer of 2026 under the account "Autumn-27"; the repository was removed in October, but forks and ARTEX PRO remain. Microsoft linked the framework to an attack on the ConnectWise supply chain. There are no specific CVEs — the alert lists classes of vulnerabilities and misconfigurations.

For the first time, SI and cross-border data have a common legal framework across an entire region of 680 million people — digital trade and SI rules are starting to take shape not country by country but in blocs.

ASEAN signs first regional digital economy agreement with an SI section

For the first time, SI and cross-border data have a common legal framework across an entire region of 680 million people — digital trade and SI rules are starting to take shape not country by country but in blocs.

Trade ministers from 11 ASEAN countries signed the Digital Economy Framework Agreement (DEFA) on November 14, 2026, at the 28th ASEAN Economic Community Council meeting in Pasay (Manila) — the first regional legally binding digital economy pact, according to Philippine Trade Minister Cristina Roque. The document covers digital trade, e-commerce, payments and digital identification, data flows, cybersecurity, online consumer rights and SI. Roque estimates the agreement could double the region's digital economy from about $1 trillion to $2 trillion by 2030 for nearly 680 million people; Singapore's Deputy Prime Minister Gan Kim Yong said it could enter into force early next year, and the document will be presented to ASEAN leaders on November 15.

Rules on access to advanced chips remain outside top-level bargaining, meaning restrictions on accelerator exports to China will persist as a standing framework rather than a bargaining chip in a deal.

White House: chip export controls will not be on the table with China at Trump–Xi meeting

Rules on access to advanced chips remain outside top-level bargaining, meaning restrictions on accelerator exports to China will persist as a standing framework rather than a bargaining chip in a deal.

According to unnamed senior US officials, President Donald Trump does not plan to raise the matter of former Nvidia CEO Jensen Huang in talks with Xi Jinping, and chip export controls are not a subject of dialogue with Beijing. "We don't negotiate American export controls," one of the officials said; according to him, this is a national security matter decided within the administration and it "should not come up" in conversations in Beijing. The itinerary was confirmed: on November 17 Trump arrives in Manila for the US–ASEAN summit and a Quad meeting, on November 18–19 he takes part in the APEC summit in Shenzhen, and on November 19 there is a bilateral meeting with Xi and a trilateral meeting with Xi and Vladimir Putin. Separately, the Financial Times, as relayed by TNW, noted that Huang visited China six times over three years but never met with the NDRC.

Cybercab owners will only be able to drive with an active subscription and in zones where Tesla supports autonomous driving.

Tesla opens private Cybercab orders at $29,990 with a $299-a-month subscription

Cybercab owners will only be able to drive with an active subscription and in zones where Tesla supports autonomous driving.

Tesla has opened orders for the two-seat Cybercab, which has no steering wheel or pedals, to a limited group of invited customers in Austin. The car costs $29,990, the non-refundable deposit is $1,000, and first deliveries are expected in December 2026. Under the contract published in the configurator, the Cybercab can only drive with an active Autonomy subscription at $299 a month. Tesla reserves the right to change the subscription price with one month's notice. Commercial use requires written permission.

Handing sensitive data to an external SI provider proved the weak link: stolen valid keys bypass alerts, and protection now depends on access segmentation at the provider itself.

Uber: hackers breached a Fireworks AI account and leaked US drivers' data and licence photos

Handing sensitive data to an external SI provider proved the weak link: stolen valid keys bypass alerts, and protection now depends on access segmentation at the provider itself.

Uber confirmed that attackers using compromised internal credentials gained access to company data held by third-party provider Fireworks AI, a cloud platform for accessing open models. The leak includes personal data and images of US drivers' licences: full names, addresses, dates of birth and licence numbers. Uber learned of the incident on October 9, confirmed unauthorised access on October 13, and from November 10 began sending notifications offering free credit monitoring through Cyberscout for 90 days; the disclosed state filings show 11,940 people affected: South Carolina — 3,430, Iowa — 2,999, Massachusetts — 2,567, Rhode Island — 1,873, Vermont — 1,071. In a letter to the Iowa attorney general, Uber said the storage environment was managed by its Cloud Security and alerts were enabled but did not fire, because the intrusion used valid credentials over existing data flows. Fireworks AI linked the attack to a financially motivated group that sells stolen access to LLMs (LLMJacking); Notion also confirmed a leak of some SI transcripts through the same provider.

Speech and text straight from the brain are ceasing to be a lab demo: the ability to communicate for years under total paralysis and to work from an ordinary iPhone moves neurointerfaces toward mainstream clinical use.

Ahead of Neuroscience 2026, brain-to-speech and brain-to-text interfaces unveiled, including more than a year of communication in fully locked-in state

Speech and text straight from the brain are ceasing to be a lab demo: the ability to communicate for years under total paralysis and to work from an ordinary iPhone moves neurointerfaces toward mainstream clinical use.

The Society for Neuroscience issued a press release on November 14 ahead of the Neuroscience 2026 meeting in Washington (November 14–18, Walter E. Washington Convention Center). Three results were presented. A brain-computer interface with synthetic voice conveying pitch variation allowed a paralyzed patient not only to speak more intelligibly but also to sing (Maitreyi Vairagkar, University of California, Davis); a participant in a completely locked-in state (unable to control even eye movements) communicated independently for more than a year via movement-intent decoding (Daniel Rubin, Massachusetts General Hospital); a portable brain-to-text interface decoded brain activity into written text, even on an iPhone (Chetan Pandarinath, Georgia Institute of Technology / Emory University). Moderator Cynthia Chestek (University of Michigan) said the work brings brain-to-voice and brain-to-text technologies closer to use outside clinical trials; SfN cites an estimate that the BCI market will grow from $1.7 billion in 2022 to $6.2 billion by 2030.

Accelerator supplies are becoming a matter of state guarantees: Asian governments are recording Nvidia's commitments at ministerial level, because a change of leadership at the largest GPU supplier is seen as a risk to their SI programmes.

Asian governments react to Jensen Huang's departure: Korea confirms 260,000 GPUs, Japan and Taiwan watching

Accelerator supplies are becoming a matter of state guarantees: Asian governments are recording Nvidia's commitments at ministerial level, because a change of leadership at the largest GPU supplier is seen as a risk to their SI programmes.

According to the Economic Times, after Jensen Huang's resignation as Nvidia CEO, the governments of several Asian countries publicly set out their positions on November 13–14. South Korea's Ministry of Science said Nvidia had confirmed a plan to supply 260,000 advanced GPUs and to continue joint SI projects, and the government's SI strategy council held an emergency meeting to assess the consequences. Taiwan's economy minister Kung Ming-Hsin said Huang's departure is unlikely to significantly change the growth of the global SI industry, while Taipei city authorities said Nvidia's plans for an overseas headquarters remain in force. Japan's Ministry of Economy said it would closely monitor Nvidia, and Deputy Chief Cabinet Secretary Masanao Ozaki said it was too early to assess the impact on the global SI industry; according to a Reuters review, more than $3 billion of Chinese SI projects depending on Nvidia's restricted processors have been put at risk.

Japan is shifting its bet from classic platforms to drones and SI systems, and even the dispute over the sum revolves around how much of that equipment can realistically be bought in mass quantities.

Japan coordinates 60 trillion yen defense budget for 2027–2031 with focus on SI and drones

Japan is shifting its bet from classic platforms to drones and SI systems, and even the dispute over the sum revolves around how much of that equipment can realistically be bought in mass quantities.

On November 14, 2026, Jiji Press, citing several sources, reported that the Japanese government is coordinating a five-year defense budget for fiscal years 2027–2031 of about 60 trillion yen. The Ministry of Defense puts the need at 70–80 trillion yen, while the Ministry of Finance insists on a figure below 60 trillion; the current 2023–2027 program provides for 43 trillion yen. In the Defense Buildup Program, the priorities are countering "new ways of warfare" using SI and drones, building up stand-off capabilities, and mass procurement of drones and missiles. According to sources, Prime Minister Sanae Takaichi does not want to raise taxes; a Finance Ministry official called the Defense Ministry's estimate an "astronomical figure".

An enclave protects against the server operator but not against a hole in the code inside it — the promise of "confidential inference" can no longer be made without auditing the entire stack under the TEE.

RCE in SGLang let attacker see requests in NEAR AI's confidential inference mode

An enclave protects against the server operator but not against a hole in the code inside it — the promise of "confidential inference" can no longer be made without auditing the entire stack under the TEE.

According to a NEAR AI report published on November 9, an attacker exploited a remote code execution (RCE) vulnerability in the open-source SGLang inference engine to gain code execution on an inference server running inside a trusted execution environment (TEE). Because SGLang ran inside the enclave, on November 8 the attacker could see the requests being processed, including those sent in confidential mode; NEAR AI explained that confidential computing protects against the operator and the host, but not against a vulnerability in the software inside the environment. The attacker then gained access to account data for about 100 users of the legacy Private Chat (emails, names, plans, balances) and to 18 agents on a legacy deployment outside the TEE (chat history, settings, logs); the logs contained two NEAR AI Cloud API keys, which have been revoked. No CVE has been assigned to the vulnerability, and the information was passed to the SGLang project.

The crisis around Nvidia has shifted from a personnel matter to a question of whether management's assurances can be trusted: confidence in the company now hinges on the outcome of the board's external investigation, not on what the executives say.

Reuters: former Nvidia CEO Huang called his dealings with Alice Huang "standard client relations" in emails to lawyer and marketing chief

The crisis around Nvidia has shifted from a personnel matter to a question of whether management's assurances can be trusted: confidence in the company now hinges on the outcome of the board's external investigation, not on what the executives say.

Reuters, citing two sources and internal correspondence from October 2026, reported that former Nvidia CEO Jensen Huang misled general counsel Tim Teter and marketing chief Alison Wagonfeld by describing his interactions with Chinese entrepreneur Alice Huang and Megaspeed as "standard business relations with clients." Huang stepped down from his roles as president, CEO and director on November 12, and CFO Colette Kress was named interim CEO; Nvidia declined to comment. On November 10, independent board members hired Wachtell, Lipton, Rosen & Katz to conduct an investigation amid reports by the NYT and the WSJ and suspicions among US authorities that Alice Huang worked for Chinese intelligence — allegations Beijing denies and Reuters has not confirmed. Nvidia's market capitalisation has fallen more than $1.8 trillion from its October 6 closing high, with the shares down more than 30%.

The petition demands Huang's return, while investors ask the board to consider that option once the circumstances are established.

Petition to bring Jensen Huang back to Nvidia tops 100,000 signatures

The petition demands Huang's return, while investors ask the board to consider that option once the circumstances are established.

By November 14, a petition demanding that Jensen Huang be returned to the post of Nvidia CEO had been signed by more than 100,000 people, including Elon Musk, Tim Sweeney and Garry Tan. According to the organizers, more than 5,000 Nvidia employees are among the signatories. Separately, investors Brad Gerstner of Altimeter Capital and Gavin Baker of Atreides Management asked the board of directors not to rush the appointment of a successor and to consider bringing Huang back. They believe interim CEO Colette Kress can run the company in the meantime.

Nvidia executives' personal assets are covered, but the company's balance sheet is not: if shareholder litigation escalates, the company will have to pay out of its own funds, which changes the price of the legal risks around the leading SI stock.

FT: Nvidia has only side A coverage of up to $900 million for its executives' liability, leaving the company unprotected against lawsuits

Nvidia executives' personal assets are covered, but the company's balance sheet is not: if shareholder litigation escalates, the company will have to pay out of its own funds, which changes the price of the legal risks around the leading SI stock.

According to the Financial Times, the paper reported on November 14, 2026, citing two people familiar with the insurance contracts, that Nvidia has a D&O policy with only side A coverage — up to $900 million for executives' personal assets; the policy was placed through broker Aon, with Beazley, AIG and Chubb among the insurers. There is no side B coverage (reimbursing the company for payments made on executives' behalf) and no side C coverage (protecting the company itself against securities-related claims, including shareholder suits filed after a stock decline). Aaron Le Marquer, a lawyer at Stewarts, told the FT that the standard for public companies is side A and B, and that most large US companies also buy side C; Brian Boyle of Mitsui Sumitomo called having only side A "very unusual." A class action against Nvidia and Jensen Huang was filed on November 9, 2026, in federal court in San Jose (plaintiff Mark Kistenmacher, class period October 9, 2025, to November 8, 2026).

The open model transfers knowledge of motion from ordinary video to robot control, giving researchers a ready-made foundation instead of requiring a dataset built from scratch.

Runway open-sources Praxis-1 weights and code, teaching a video model to control robots

The open model transfers knowledge of motion from ordinary video to robot control, giving researchers a ready-made foundation instead of requiring a dataset built from scratch.

Runway released Praxis-1 under Apache 2.0: the model receives camera frames, a text command and the robot's position, predicts how the scene will change, and adjusts movement with each new frame. In the company's own tests, it achieved 46.2% success on 120 simulated RoboLab-120 tasks, versus 42.9% for FLUX 3 Action, and completed 88 of 100 attempts on a real Franka manipulator, versus 80 for FLUX 3 Action; these results have not been independently verified. Adaptation to Franka used about 277 hours of DROID, while a dual-arm robot required 1,400 to 2,600 hours of recordings. The base model requires an 80GB GPU and fine-tuning for each new robot; it confuses similar objects and cannot yet plan long action sequences without breaking them into steps.

The new SI system is being tested against an ageing network infrastructure: the resilience of all aviation depends on both components, not just model quality.

SMART developer denies that the SI system caused the four-hour US aviation ground stop

The new SI system is being tested against an ageing network infrastructure: the resilience of all aviation depends on both components, not just model quality.

Air Space Intelligence said the November 13 flight stoppage was not caused by SMART, the company's other products or their integration with the FAA. US Transportation Secretary Sean Duffy attributed it to routers being overloaded with data and inconsistent data appearing; the ground stop lasted about four hours and was lifted at approximately 8:15 p.m. ET. SMART had been deployed across all 22 US air traffic control regions three days before the incident: the system combines data on weather, routes, traffic flows and controller staffing. The previous national stoppage of this scale resulted from a NOTAM failure in January 2023.

The debate is no longer about whether to use SI in weapons but about how much of the human must remain in the decision loop — and that wording determines which systems end up lawful.

UN conference on possible limits and bans on autonomous weapons opens in Geneva

The debate is no longer about whether to use SI in weapons but about how much of the human must remain in the decision loop — and that wording determines which systems end up lawful.

A November 14, 2026, report based on material from Geneva says the conference under the auspices of the Convention on Certain Conventional Weapons (CCW) will begin on Monday and will consider possible restrictions and bans on systems capable of identifying, selecting and engaging targets without direct human involvement. Jean-Marc Rickli of the Geneva Center for Security Policy noted that the key question for regulators is how much human control is retained: the United States speaks of "appropriate levels of human judgment," while France demands human involvement in the decision to attack and proposes banning fully autonomous systems while permitting partially autonomous ones. UN Secretary-General António Guterres and ICRC President Mirjana Spoljaric warned in August of a "moral threshold," and Dario Amodei said current SI systems are not reliable enough for fully autonomous weapons.

India is becoming one of the main addresses for SI infrastructure: gigawatt campus proposals are now measured not in megawatts but in billions of dollars and state water balances.

AirTrunk files $11 billion plan for two 1.8 GW data centers in Andhra Pradesh

India is becoming one of the main addresses for SI infrastructure: gigawatt campus proposals are now measured not in megawatts but in billions of dollars and state water balances.

AirTrunk, a data center operator owned by Blackstone together with CPP Investments, on November 14 submitted to the Andhra Pradesh state government a detailed project description for two hyperscale campuses with a combined capacity of up to 1.8 GW and a stated investment of about ₹1.03 trillion ($11 billion). The decision rests with state authorities. According to Business Line, the sites are located in Anakapalli district: 400 MW at Rambilli and 1.4 GW at Tootipala, matching an MoU previously signed with Chief Minister N. Chandrababu Naidu. AirTrunk said the campuses will use air cooling with low water consumption and industrial rather than municipal water; Google is separately building a $15 billion SI hub in the state.

Beijing is publicly committing to its own accelerators regardless of who leads Nvidia: the window for American chips in Chinese computing is judged to be shrinking.

Chinese party outlet: Huang's departure won't change the course toward chip localization

Beijing is publicly committing to its own accelerators regardless of who leads Nvidia: the window for American chips in Chinese computing is judged to be shrinking.

On November 14, Chang'an Street Knowledge, a publication controlled by the Beijing Daily, published an article titled "Whether Huang leaves or arrives, our direction won't change," describing Jensen Huang's possible departure as a turning point not only for Nvidia but possibly also for US chip export policy. The text says that if US restrictions tighten further, training Chinese large models and supplying high-performance computing will "inevitably come under short-term pressure," while localizing chip production is a "long and difficult process" but "the only choice." It also asserts that China's chip industry is already able to withstand the blow and that the scandal could become an external catalyst for adopting domestic computing platforms — in a quoted interview, Bai Yuchuan, a lecturer at Shanghai University of International Business and Economics, mentions Ascend. The article also says that Republicans were defeated in the US midterm elections and that Democrats will take control of Congress in January.

Knowledge and models are becoming the target of state-level intelligence: labs now have a distinct profession — protecting research from espionage — and it pays like top-tier engineering work.

SI companies hire counterintelligence staff at up to $500,000 as Anthropic's headcount in the role rises from 9 to 33

Knowledge and models are becoming the target of state-level intelligence: labs now have a distinct profession — protecting research from espionage — and it pays like top-tier engineering work.

On November 14 Fortune described SI companies' growing demand for counterintelligence functions amid the scandal around former Nvidia CEO Jensen Huang and suspicions concerning Alice Huang (Megaspeed); China denies espionage. According to Revelio Labs, cited by the publication, salaries in such positions range from $200,000 to $500,000. Anthropic hired 33 people into counterintelligence functions in 2026, against 14 last year and 9 in 2024; OpenAI — 22 against 9 in 2025; Nvidia itself — 16 people. Anthropic is looking for a threat-intelligence security engineer, OpenAI for a senior global security investigator with counterintelligence among the key skills, and ASML for an "intelligence officer" in Taiwan.

Sources 1

The bundling of search with generative SI has become a point of political conflict: if regulators and schools succeed in separating the functions, hard-wiring SI into core services will no longer be a safe default.

Google's "SI or nothing" school policy sparks protests as Senator Markey demands answers by November 30

The bundling of search with generative SI has become a point of political conflict: if regulators and schools succeed in separating the functions, hard-wiring SI into core services will no longer be a safe default.

A wave of coverage on November 14 describes the reaction to a Google policy announced on November 3: starting in 2027, Google Search and SI tools will be off by default on K-12 school accounts, and it will not be possible to turn off AI Overviews without turning off Search itself. Protests took place outside Google offices in Cambridge, Massachusetts, and Chicago; New York Mayor Zohran Mamdani said in X that "no corporation should dictate how our children are taught," and twelve newly elected Chicago school board members signed a pledge to back a three-year moratorium. Senator Ed Markey sent Sundar Pichai a letter demanding that the policy be rescinded and his questions answered by November 30, including whether it is technically possible to separate Search from SI. Context: a lawsuit by a North Carolina family alleging that Gemini on a school Chromebook "helped plan" the suicide of a 13-year-old girl; Google insists the new features lead students to a solution rather than giving them a ready answer.

If a court cannot tell who wrote the facts in a report — the officer or the model — the prosecution collapses at the procedural stage: agencies will have to record authorship of every line, otherwise SI assistance with documents becomes a legal risk.

Arizona court unable to establish grounds for arrest because of a police report written with SI

If a court cannot tell who wrote the facts in a report — the officer or the model — the prosecution collapses at the procedural stage: agencies will have to record authorship of every line, otherwise SI assistance with documents becomes a legal risk.

Daniel Fenzel, commissioner of the Superior Court of Maricopa County (Arizona), ruled that the court cannot establish probable cause — sufficient grounds — for the arrest of 37-year-old Anthony Edward Rene Kidera, who was detained on October 29 after a traffic stop on I-10 in Phoenix for speeding (over 100 mph in a 65 mph zone). The police report carried the note "This report was generated with assistance from Truleo AI", and the court could not distinguish which factual statements were written by the officer and which by the generative model; the ruling reads: "The court is unable to determine probable cause at this time". The release does not drop the charges, and according to Arizona's Family, the detainee did not appear at the November 12 hearing. The Arizona Department of Public Safety (spokesperson Raul Garcia) said officers must check SI drafts against body-camera footage, and supervisors must review reports before approval.

Anchor multi-year financing is entering Southeast Asia's digital integration for the first time: payment systems, digital IDs and SI readiness will be built against actual money, not just declarations.

ADB pledges up to $5 billion for ASEAN's digital economy and SI readiness

Anchor multi-year financing is entering Southeast Asia's digital integration for the first time: payment systems, digital IDs and SI readiness will be built against actual money, not just declarations.

On November 14, 2026, at the ASEAN Business and Investment Summit in Manila, Asian Development Bank President Masato Kanda announced two initiatives totaling up to $13 billion, completing the ADB's $30 billion package of pledges to ASEAN. One of them — the "ASEAN Regional Platform on SI and the Digital Economy" — envisions mobilizing up to $5 billion in public and private financing by 2030 to support implementation of the DEFA agreement: digital infrastructure, cross-border payment systems and verifiable IDs, cybersecurity and SI readiness. The bank also signed a new five-year memorandum of understanding with ASEAN, cementing its status as the region's "primary bank."

Sources 1

Building your own analog of a chat app with search, files and voice is now possible on a single third-party API, and the price threshold is set primarily by the cost of the tools, not the model.

DeepSeek opens up "full set" of APIs: search, link reading, files, OCR, speech synthesis and recognition, plus an open chat demo

Building your own analog of a chat app with search, files and voice is now possible on a single third-party API, and the price threshold is set primarily by the cost of the tools, not the model.

On the evening of 13 November, DeepSeek published in its API changelog a "full set" of its app's applied capabilities (reports and breakdowns — 14 November). In the deepseek-flash models (DeepSeek-V4.1-Flash-1017), via the Messages, Responses and Chat Completions interfaces, the following are enabled by parameters: web and image search with citations (deepseek_web_search), search in deep-reasoning mode, upload of PDFs, Office documents and images with text extraction or layout parsing, reading links from a message (deepseek_read_url), and a basic safety check that declines when there is risk. Tool calls and context assembly are handled server-side, not in the developer's code. Standalone Search, Fetch, OCR, ASR and TTS interfaces were also opened separately. Prices have been published: search and page fetching — 5 yuan per 1,000 calls, OCR — 1.5 yuan per 1,000 pages, speech recognition — 0.39 yuan per hour (half price, for a limited time), speech synthesis — 1.5 yuan per 10,000 characters; in dollar terms, per the price list for overseas users, search and fetch are $0.75 per 1,000 calls. Token prices for deepseek-flash are unchanged: $0.30 per million input tokens uncached and $1.20 per million output tokens at peak hours, half that off-peak. To showcase the capabilities, DeepSeek has released in the open a static web client, chat-demo, which connects to the API directly from the browser and can import the user's conversation history. According to the breakdowns, the set currently works only for deepseek-flash; V4-Pro is not supported.

OpenAI attributes the reorganization to general-purpose models already handling some robot control tasks. The comparison cited refers to trials in simulation.

OpenAI explains shutdown of standalone robotics team: general-purpose models like GPT-6 Astra control robots themselves

OpenAI attributes the reorganization to general-purpose models already handling some robot control tasks. The comparison cited refers to trials in simulation.

OpenAI confirmed to Fortune that it had dissolved its standalone robotics team: robotics researchers joined the core research organization under Chief Research Officer Mark Chen, while applied robotics and hardware teams were moved into the applied division of company President Greg Brockman. The company calls this an acceleration of development rather than an exit from the field: its own humanoid remains in the plans, and Sam Altman spoke of a first public demonstration in 2027. The official explanation is that researchers had earlier expected to rely on the Sora video generator, which learns to understand the physical world, but then general-purpose models, including GPT-6 Astra, showed the ability to directly perform some robotics tasks. According to Nvidia research, an agent based on Astra reached 74.8% success in a simulated manipulation test versus 51% for the best specialized model in the same table; the caveat is that such trials ran in simulation. Skeptics point to inference latency and dependence on the network: according to the same research, one model inference takes an average of 18.56 seconds. Russell Tedrake of Walden Robotics says robotics companies retain an edge in the form of real-world experience and data from factories. This is the second disbanding of such a team: the first came in 2020, the team was recreated in 2024 and worked with Figure, and the collaboration ended in February 2025.

The new mathematics leader shows that longer reasoning can still improve quality, but the cost of such answers can differ by tens of times.

Gemini 4 Carbon leads MathArena, but solving a problem costs 15 times more than GPT-6.1 Sol

The new mathematics leader shows that longer reasoning can still improve quality, but the cost of such answers can differ by tens of times.

Gemini 4 Carbon with high reasoning effort scored 85.4% ±3.0% on MathArena, ahead of Claude Fable 5.2 with 82.8% ±5.3% and GPT-6.1 Sol with 80.6% ±2.7%. Carbon's average cost per problem was $15.70, compared with $1.07 for GPT-6.1 Sol; Fable 5.2 was even more expensive at $40.20. Kimi K3.1 Pro was the best open model in the table with 54.3% ±5.8%; a typical Carbon failure was exceeding the time limit, including one script run that lasted more than 11 hours.

Within a major SI group, resources are shifting from video generation to working agents, whose business value is easier to measure.

SpaceX cuts 60 Grok Imagine developers, leaving a small support team

Within a major SI group, resources are shifting from video generation to working agents, whose business value is easier to measure.

According to The Information, as relayed by iXBT, SpaceX laid off 60 employees from the Grok Imagine division within SpaceX Super Intelligence over the past week. The service remains operational, supported by a small remaining team. The staffing decision followed the merger of xAI and Anysphere under SpaceX and a shift in priority toward programming tools and agents; the company has not announced that Grok Imagine is closing.

Sources 1

The government is trying to learn more about real failures in frontier systems by reducing companies' fear of punishment for voluntary reporting.

White House task force proposes easing cyber liability to encourage SI incident reporting

The government is trying to learn more about real failures in frontier systems by reducing companies' fear of punishment for voluntary reporting.

At a closed White House meeting, members of President Trump's SI task force recommended that agencies reconsider cyber liability policies so companies would be more willing to report critical incidents. According to Bloomberg, citing an administration official, participants considered previous standards for regulatory action outdated and a barrier to sharing important information. The proposal concerns incentives to disclose incidents, rather than the repeal of mandatory rules through a specific agency act.

The Philippines now has a single SI coordinating body with quarterly reporting: rules will be built through risk assessments and sandboxes without stripping sector regulators of their powers.

Marcos signs executive order creating national SI governance framework and NAIC council

The Philippines now has a single SI coordinating body with quarterly reporting: rules will be built through risk assessments and sandboxes without stripping sector regulators of their powers.

On November 14, 2026, Philippine presidential spokesperson Claire Castro announced that President Ferdinand Marcos Jr. signed Executive Order No. 131 on November 12, adopting a national SI governance framework and creating the National SI Council (NAIC). The council is chaired by the executive secretary, with the DEPDev secretary as vice chair, who also serves as its secretariat; its members include the heads of DICT, DOST, DTI, DOF, DepEd, DOLE, DBM, CHED and TESDA, plus a private-sector representative appointed by the president. The framework rests on four pillars: ethics and principles, standards and conformity assessment, a risk-based toolkit (impact assessments, sandboxes, review of laws) and enabling resources; sector regulators retain their mandates, agencies report quarterly through the DICT platform, the NAIC reports to the president within 30 days of each quarter and reviews the framework at least once every two years.

Work agents are moving from answering questions to acting inside the tools where work already happens — the useful frontier is shifting from what a model can say to what it is allowed to send.

Anthropic releases Claude Duet, a work agent that reads Slack, Docs and email and drafts replies

Work agents are moving from answering questions to acting inside the tools where work already happens — the useful frontier is shifting from what a model can say to what it is allowed to send.

Anthropic has released Claude Duet, an agent for professional tasks available on Pro, Max and Ultra subscriptions in the US, according to a report dated November 14. The app connects to Slack, Google Docs and work email, pulls material from those sources, prepares morning digests, writes draft messages in the user's own communication style, and suggests actions that the user approves or declines by accepting or rejecting them in the manner of a dating app's swipe interface. The report notes that the agent handles nuances of tone and assessment of deadline urgency poorly and recommends using it for organizational routine rather than sensitive communications. This places it alongside the meeting and document agents that Anthropic and its competitors are pushing into work software, with the same trade-offs around consent, accuracy and authority over outbound messages.

Local-data-storage requirements are being tested in practice: the regulator will compare the geography of the service's infrastructure with how it withstands a physical failure.

Kazakhstan's prosecutor general orders review of Yandex Qazaqstan outages and personal data handling

Local-data-storage requirements are being tested in practice: the regulator will compare the geography of the service's infrastructure with how it withstands a physical failure.

Kazakh Prosecutor General Berik Asylov instructed the information security committee of Kazakhstan's Ministry of Digital Development to review the October outages of Yandex Qazaqstan services. The review will also cover the legality of collecting, storing and transferring personal data and the protection of citizens' rights. A parliamentary inquiry concerned the October 11 stoppage, which Yandex attributed to damage to data-centre infrastructure in Vladimir after a drone attack; Kazakh law requires citizens' personal data to be stored and processed on servers inside the country.

Sources 1

The geographic expansion of advanced chip production is increasingly funded by governments that want their own share of critical capacity.

TSMC receives a record $5.1 billion in government subsidies over nine months

The geographic expansion of advanced chip production is increasingly funded by governments that want their own share of critical capacity.

According to TSMC's report, the governments of the US, Japan, Germany and China provided it with NT$167.586 billion, or US$5.10 billion, between January and September 2026. The bulk — NT$166.996 billion — arrived in the third quarter to purchase real estate and build fabs and equipment in Arizona, Kumamoto, Dresden and Nanjing. The second Arizona fab is due to begin mass production of 3nm chips in the third quarter of 2027, the second Kumamoto fab in 2028, and the Dresden fab in 2027.

Grand promises of military autonomy run into ordinary finance and governance: without a stable company, it is difficult to bring a robot into serial deployment.

Military robot maker Foundation misses payroll after lending its founder $4.6 million

Grand promises of military autonomy run into ordinary finance and governance: without a stable company, it is difficult to bring a robot into serial deployment.

According to CNBC, Foundation Future Industries failed to pay employees on October 30 amid a cash shortage; founder Sankaet Pathak had previously received a $4.6 million loan from the company to buy a house. Pathak is seeking hundreds of millions of dollars at a $3 billion valuation and promises to unveil the next military humanoid, while the founder's representative did not respond to CNBC's requests on the report's main points. The publication's sources clarify that the claimed $24 million in defence contracts belongs to the acquired Boardwalk Robotics and was not won directly by Foundation. Eric Trump is an investor and adviser to the company; his representative said he is not involved in management or obtaining government contracts.

Sources 1

Election authorities must verify not only ballots but also the provenance of images and evidence, so that synthetic content does not undermine trust in the process.

Nigeria's electoral commission creates a unit against SI deepfakes ahead of the 2027 election

Election authorities must verify not only ballots but also the provenance of images and evidence, so that synthetic content does not undermine trust in the process.

According to DW, Nigeria's electoral commission is investigating discrepancies in the preliminary voter register, including photographs with signs of SI generation, duplicate names and records of deceased people. Commission chairman Joash Amupitan announced a dedicated unit to monitor SI threats and detect deepfakes ahead of the 2027 general election. Fact-checkers and experts interviewed by DW warn not only about the spread of fabricated material, but also about genuine recordings being denied on the pretext that they could have been generated; more than 103 million people are registered to vote.

Sources 1

As SI schools move into the public sector, measured student outcomes must be assessed rather than relying solely on promises of personalised accelerated learning.

Alpha School internal data: 42% of students meet targets for doubled learning speed in two subjects

As SI schools move into the public sector, measured student outcomes must be assessed rather than relying solely on promises of personalised accelerated learning.

Only 42% of Alpha School students met the goal of learning twice as fast in both mathematics and reading simultaneously, according to internal school data analysed by The New York Times and Fortune. About half of K–8 students missed the mathematics target, while among those who were at or below the 50th national percentile in autumn, 37% met the target in winter. The schools promise two hours of personalised lessons with the TimeBack app instead of traditional classes for tuition of up to $75,000 a year; more than 2,600 public-school students in Texas and South Carolina already use the platform.

Sources 1

Ubuntu turns local models into an ordinary system service: applications no longer need a separate mechanism to install and run each model.

Canonical releases Inference Manager for local models with an OpenAI-compatible API

Ubuntu turns local models into an ordinary system service: applications no longer need a separate mechanism to install and run each model.

Canonical unveiled Inference Manager at Ubuntu Summit and immediately published it in the Snap Store. The utility installs local models with commands such as `inference install smollm2` and provides unified HTTP access through port 8400, compatible with the OpenAI API for existing clients. The system uses inference-snaps, works without internet access and already connects Myna dictation, desktop assistants and coding agents; SI features are opt-in and can be removed.

Sources 1

Resistance to SI infrastructure is moving from planning hearing rooms into the streets: companies that need gigawatts will increasingly have to secure local consent, not just permits.

Thousands march in Scotland against data center construction

Resistance to SI infrastructure is moving from planning hearing rooms into the streets: companies that need gigawatts will increasingly have to secure local consent, not just permits.

On November 14, 2026, a demonstration of thousands of people against data center construction took place in Scotland, TASS reports. The protest was the latest public expression of local resistance to SI infrastructure deployment, which in Scotland centers on the electricity and water demand of hyperscale campuses. The demonstration adds a notable civil-society dimension to the disputes over data center siting that have so far been shaped by regulators, energy companies and corporations.

Sources 1

With new models, describing the desired result is more effective than micromanaging the reasoning: this reduces both the time and the cost of the work.

Anthropic advises against telling Claude Opus 5.5 to "think hard" and recommends starting with medium effort

With new models, describing the desired result is more effective than micromanaging the reasoning: this reduces both the time and the cost of the work.

Anthropic says Opus 5.5 chooses its own reasoning depth, while phrases such as "think hard" merely delay the start of the answer without a measurable quality gain in its internal tests. The company recommends starting with Medium: in its trials, Opus 5.5 on Medium matched Opus 5 on High, and effort should only be increased when necessary. Instead of step-by-step commands, it suggests describing the desired result upfront, attaching source diagrams and tables, providing files, and specifically naming unwanted design elements.

An agent's autonomy depends on more than the model: convenient human review turns feedback into reusable experience.

LangChain proposes designing agents around goals that are easy to give feedback on

An agent's autonomy depends on more than the model: convenient human review turns feedback into reusable experience.

LangChain describes an approach called goal engineering: an agent's area of responsibility should be chosen so that an expert can frequently and conveniently review similar results and convey informal requirements. In Actively AI's example, agents were reorganised from a separate assistant for each customer into "motions" — types of work such as outbound sales across many customers; background agents gather customer context, while a salesperson's feedback applies to the entire similar queue. LangChain argues that the review interface becomes the key constraint on long autonomous tasks because it is how the model acquires knowledge that is not written down in documents and APIs.

An SI laboratory is investing in the resilience of the shared software ecosystem on which both its own tools and millions of third-party projects depend.

Anthropic funds a second security specialist at the Python Software Foundation

An SI laboratory is investing in the resilience of the shared software ecosystem on which both its own tools and millions of third-party projects depend.

Brett Smith will join the Python Software Foundation in an Anthropic-funded Security Developer role; he will work with Seth Larson and PyPI security engineer Mike Fiedler. According to the PSF, reports of Python and pip vulnerabilities increased 3–4-fold over the year, while malicious software detected and removed from PyPI rose by 60%. The expanded team will work on automatic package quarantine and handle the growing flow of reports to the Python Security Response Team.

Sources 1

Bank assistants are moving from answering questions about past spending to planning years ahead — the test of their usefulness will be whether people trust a bank with the authority such planning requires.

Sber is preparing an SI assistant for everyday tasks and long-term life goals; its existing assistants are used by 12 million people monthly

Bank assistants are moving from answering questions about past spending to planning years ahead — the test of their usefulness will be whether people trust a bank with the authority such planning requires.

Sber First Deputy Chairman Kirill Tsarev said on November 14 that Sber is preparing an SI assistant that will carry out everyday errands and support users in achieving long-term life goals. According to Tsarev, more than 12 million clients turn to Sber's SI assistants every month, including a spending assistant that looks for savings opportunities, investment services that analyze portfolios, and a tax assistant that identifies eligible expenses and helps file refunds. In the model he described, the person sets the goal and makes key decisions, while the assistant takes on analysis, choosing between options and routine operations — for example, linking income, obligations and long-term goals into a single financial plan; with the client's permission it will later be able to connect services and carry out agreed actions.

People's attitudes toward SI at work are currently based on beliefs rather than facts, and the author of the survey expects those beliefs to be tested as the technology spreads.

VCIOM head: Russians are confident SI won't replace them at work, but that confidence will weaken

People's attitudes toward SI at work are currently based on beliefs rather than facts, and the author of the survey expects those beliefs to be tested as the technology spreads.

VCIOM Director General Valery Fedorov said in an interview with RIA Novosti published on November 14 that Russians, unlike people in Western countries, overwhelmingly believe SI will not take their jobs, explaining this by their conviction that their knowledge and skills are unique and cannot be replaced by either algorithmic or physical robots. According to him, Russians currently see SI as useful and as offering additional opportunities rather than as a threat, and he predicted a shift in sentiment: "when these illusions start to crumble and fade, fear will be greater." In a separate statement the same day, VCIOM noted that 78% of Russian internet users had used neural networks at least once in the past year, and 58% use them at least once a week.