Show only each day's banner story

SI helps find two critical PHP vulnerabilities: emergency updates released, some government sites taken offline during patching

SI can now turn an old bug into a working exploit in hours, so fixes have to come just as fast; this time defenders got there a little earlier.

On October 21, PHP released updates 8.5.12, 8.4.27, 8.3.36 and 8.2.35 (plus fixes for the unsupported 7.4, 8.0 and 8.1 branches) for two memory-handling bugs, CVE-2026-109618 and CVE-2026-109619 (9.2 on the CVSS scale), which allow code to be executed on a server without logging in. According to Wiz, all versions from PHP 5.4 onward are affected, and PHP runs roughly 70% of websites. Wiz names Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol as the tools used; it says SI cut the path from a bug to a working exploit chain from months to hours. According to TNW, the firm Calif found the first bug in August using GPT-5.6 Sol. According to the BBC, agencies in several countries advised organisations that had not managed to update to consider taking their sites offline: Finland temporarily shut down its public government services, the Dutch cybersecurity agency took its own site down, and the US NSA gave similar advice. According to TNW, Cloudflare switched on protective rules for all customers nine hours before disclosure, and OpenAI said it expects a sharp rise in similar findings.

Sources 3

wiz.iobbc.co.ukthenextweb.com

Claude now has an "always-on" assistant that comes to you with tasks; Meta, OpenAI and Anthropic each have such an agent.

Anthropic unveils Claude Duet, an agent that comes to you with tasks and keeps working while you're away; US waitlist only for now

Claude now has an "always-on" assistant that comes to you with tasks; Meta, OpenAI and Anthropic each have such an agent.

Anthropic announced Claude Duet as a "research preview": every morning the agent sends a short brief on connected apps, during the day it messages you only when something needs attention, takes on tasks from scheduling meetings to dashboards, financial forecasts and web apps, and, according to Anthropic, asks for confirmation before acting on the user's behalf. It runs on Opus 5.5 and is available as iOS and Android apps and through the Claude web app. According to Engadget, for now there is a waitlist for US subscribers; after launch, Pro subscribers ($20 a month) will get Duet, while OpenAI's Dots starts at $100 a month; Duet does not see passwords and sign-in tokens directly, and access to health and financial data requires explicit permission. According to CNBC, Anthropic describes Duet as a product for small businesses and professionals; Meta's Muse and OpenAI's Dots came out earlier.

Sources 3

cnbc.comtheverge.comengadget.com

A small team built a strong model by handing SI almost the whole process: a step toward SI improving SI.

Startup Recursive releases Flint, an open coding model whose pipeline, it says, was built almost entirely by SI

A small team built a strong model by handing SI almost the whole process: a step toward SI improving SI.

Flint is the first model from Richard Socher's company Recursive: open weights, 40 billion parameters. According to Socher, the company's system "Eureka Machine" automated most of the development: it devised a way to train the base model, new architectures and code for specific chips, so a small team released a best-in-class model for its size within a few months. The benchmarks were supplied by the company itself: by them, Flint beats models of similar size and architecture at programming. The Wired author ran Flint on a Mac with 48 GB of memory and wrote that it handled all his tasks, although models of this size are usually worse. Recursive has raised $650 million at a $4.65 billion valuation. Some experts fear that fully recursive self-improvement of SI could slip out of control, and there are bills in Congress to ban the approach; Recursive considers the fears exaggerated and says the process remains under human control.

Sources 1

wired.com

Meta is spending almost twice as much on SI as a year ago and is opening a new way to raise money for it — the bet on SI keeps growing.

Meta's quarterly revenue is $64.6 billion and capital spending $34.9 billion; the company creates a new non-voting share class to fund SI investment

Meta is spending almost twice as much on SI as a year ago and is opening a new way to raise money for it — the bet on SI keeps growing.

According to the third-quarter report, Meta's revenue rose 26% to $64.6 billion (expected: $63.3 billion); spending on equipment and data centers for the quarter was $34.9 billion versus $19.4 billion a year earlier; the full-year guidance is $135–145 billion. Because of roughly $10 billion in legal payments, operating income fell 37% to $13 billion, and free cash flow turned negative ($–2.1 billion). The company is introducing a non-voting Class C share: it is meant to fund SI infrastructure, acquisitions and employee compensation, with trading starting in 2027. According to TechCrunch, the Muse assistant has more than 10 million users, about a quarter use it daily, and shopping is one of the main activities; Zuckerberg said Muse monetization will start sooner than usual.

Sources 3

fortune.comtheverge.comtechcrunch.com

A ready-made open framework with SI agents let a state-linked group run mass attacks and cut the time to cloud access to hours.

Microsoft: China-linked group Storm-0940 used SI agents of the ARTEX framework to breach a ConnectWise application and gained cloud access within hours

A ready-made open framework with SI agents let a state-linked group run mass attacks and cut the time to cloud access to hours.

According to Microsoft, the group Storm-0940, which the company describes as aligned with the Chinese state, has since September used ARTEX — an open-source multi-agent penetration-testing framework that works with any SI model — to attack hundreds of organizations across industries, from aerospace and defense to finance and government. Microsoft holds the group responsible for the recently disclosed compromise of the multi-tenant ConnectWise Cloud Backup application: using stolen application credentials, the agents scanned networks, guessed passwords and abused permissions in Microsoft 365 environments without an interactive user sign-in. In several cases less than three hours passed from selecting a target to privileged cloud access. According to The Next Web, sign-in attempts were made against more than 2,250 Microsoft 365 tenants, Huntress confirmed breaches at 162, and the rate reached 2,900 sign-in attempts per minute. Microsoft stresses that no vulnerabilities in its products were used, most of the attacks failed, and the presence of an affected application does not mean data was stolen.

Sources 2

microsoft.comthenextweb.com

OpenAI's models can now be bought on all three major US clouds, though at Google only selected customers have access for now.

OpenAI models GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna and Codex appear on Google Cloud in closed preview

OpenAI's models can now be bought on all three major US clouds, though at Google only selected customers have access for now.

According to OpenAI (as relayed by The Next Web), the models and the Codex agent have been added to Google Cloud's Model Garden: a closed preview for selected enterprise customers is already running, and general availability is promised "in the coming weeks". They can be bought through Google Cloud Marketplace against existing spending commitments, on a single invoice; the models run on dedicated Google infrastructure, and Google, according to the parties, does not use prompts and responses for training. OpenAI stresses that Microsoft remains its primary cloud partner; according to TNW, Microsoft's exclusive licence has ended, and Amazon Web Services has hosted OpenAI models since April 2026.

Sources 2

thenextweb.comopenai.com

Local resistance to data centres has become a political issue, and the authorities are looking for ways to "share the benefit" with residents.

Trump proposes that data centres pay "dividends" to residents and communities to win their support

Local resistance to data centres has become a political issue, and the authorities are looking for ways to "share the benefit" with residents.

In a Truth Social post on October 21, Donald Trump wrote that companies wanting data centres to be popular should pay the community and/or its residents a "dividend"; in his words, unpopular centres become "extremely popular" once a financial deal is struck. According to Al Jazeera, polls show 71% of Americans (Gallup, March) and 72% (ABC News/Washington Post/Ipsos, October) object to data centres near them, mainly citing water and electricity use; several Republican Senate candidates have changed their earlier position and now oppose data centres. In August Trump said that objections benefit China.

Sources 2

aljazeera.comcnbc.com

A merger would put SI, robots and space under one company, but for now this is only a study of the question.

Musk: SpaceX and Tesla are studying a possible merger, but nothing is decided

A merger would put SI, robots and space under one company, but for now this is only a study of the question.

According to Al Jazeera, on SpaceX's quarterly earnings call Elon Musk said the companies have set up special committees to study a merger, and called the idea one with "a lot of merit" because of shared areas — SI, robotics, manufacturing and energy. He cautioned that treating the merger as settled would be overstating it; it would require Tesla shareholder approval. Their combined market value, by the same account, is about $3.5 trillion, which would make it the largest deal in history. For the third quarter SpaceX reported its first quarterly profit, $4.4 billion; SpaceX shares rose more than 8% in after-hours trading and Tesla's rose 1.8%.

Sources 2

aljazeera.comsemafor.com

Many agents at once can quickly find ideas for biologists in data and literature, but they still have to be tested in the lab.

OpenAI: thousands of SI agents analysed data on millions of mouse brain cells in 32 minutes and proposed new directions to test

Many agents at once can quickly find ideas for biologists in data and literature, but they still have to be tested in the lab.

According to OpenAI (as relayed by superpowerdaily; OpenAI's own page is closed to reading), OpenAI, PerturbAI and Scripps Research used agents built on GPT-Rosalind to turn an existing CRISPR dataset on the mouse brain into an atlas linking the effects of gene knockouts, publications and testable mechanisms. In a comparison of ways to organise the work, distributed agents at maximum scale cited more than 700 papers versus 11 for a single agent, and experts in a blind evaluation preferred the distributed agents' reports 75% of the time. This is an assessment of analytical breadth and usefulness, not experimental confirmation: the proposed mechanisms still need to be tested in the lab.

Sources 2

openai.comsuperpowerdaily.com

Proteins can now be "ordered" in words and they already work in the lab; this makes it easier to create enzymes for industry.

A 16-billion-parameter model, Pinal, created working proteins from text descriptions, including an enzyme and a protein for capturing CO2

Proteins can now be "ordered" in words and they already work in the lab; this makes it easier to create enzymes for industry.

According to the abstract of the Nature paper, Pinal designs functional proteins from natural-language descriptions: it first generates a structure from the text, then selects a sequence. The model was applied to four tasks — a fluorescent protein, a PET hydrolase, an alcohol dehydrogenase and the H-protein of a metabolic pathway — and in each case the intended function was observed experimentally; catalytic activity was confirmed for both enzymes. Without iterative experimental optimisation, the designed H-protein raised product yield in a multi-enzyme CO2-fixation pathway 1.7-fold compared with the native E. coli protein. Training used 1.7 billion synthetically annotated protein–text pairs.

Sources 1

nature.com

A protein design tool is becoming open and available in clouds and labs; how well its designs work in the test tube will be shown by experiments.

Nvidia and the University of Washington's Institute for Protein Design release RFD4-Proteina, an open model for designing proteins across tasks

A protein design tool is becoming open and available in clouds and labs; how well its designs work in the test tube will be shown by experiments.

RFD4-Proteina is an open model built by Nvidia together with the University of Washington's Institute for Protein Design (UW Medicine). According to Databricks, it generates both the amino-acid sequence and the full atomic structure of a protein in a single pass, and reward-guided search can steer a design toward desired properties such as thermostability or binding; there are ready-made workflows for antibodies and vaccines. According to Nvidia, on four benchmark suites (protein binding, small-molecule binding, enzyme scaffolding and DNA binding) the model showed a higher rate of successful computational designs and higher generation throughput than the baselines it was compared with. The model is available through partners including Hugging Face, Google Cloud, Databricks, Benchling, Anthropic's Claude Science and OpenAI. Genedata and Danaher will use it in an autonomous lab where experimental results guide the next design cycle.

Sources 2

databricks.comnvidianews.nvidia.com

The chip shortage comes down to substrates, packaging and testing, and Nvidia is paying up front to get more of them.

Nvidia and six manufacturers expand output of substrates, packaging and chip testing for SI; Nvidia advances KYEC $1.1 billion

The chip shortage comes down to substrates, packaging and testing, and Nvidia is paying up front to get more of them.

According to Nvidia, IBIDEN, Samsung Electro-Mechanics and Unimicron will together increase substrate capacity 2.5 times over three years (IBIDEN's plant expansion — threefold by the first quarter of 2028); Amkor will raise interposer capacity by more than 300% by the end of 2028; testing capacity will double from 2026 to 2027, and Nvidia is giving KYEC a $1.1 billion advance for new test facilities in the US (volume production from April 2027); SPIL plans to raise output of optical engines for co-packaged optics 15 times by mid-2027. Jensen Huang, according to Nvidia: "Capacity is the constraint."

Sources 1

nvidianews.nvidia.com

Investors are putting a billion into the idea that SI can be trained on far fewer examples.

One-year-old lab Flapping Airplanes raises $1 billion at a $6 billion valuation for models that need far less data

Investors are putting a billion into the idea that SI can be trained on far fewer examples.

The Series A round was led by Index Ventures and Lux Capital, with Sequoia, GV, Kleiner Perkins, Valor, Nvidia and others participating; in total the company has raised $1.18 billion including the seed round. The team is about 20–25 people, and there is no product on sale yet. The lab is betting on SI models trained on much less data than today's frontier systems; the basis is a "megakernel" that runs processes on the GPU as a single kernel. Co-founder Asher Spector says a model that is "a million times more data-efficient" would be "a million times easier to put into the economy."

Sources 1

siliconangle.com

Gestational age can be determined accurately without a sonographer — this brings prenatal care closer to places without specialists.

Google's SI reads a "blind" ultrasound by an untrained worker and gets gestational age to within 4.18 days — as good as experts

Gestational age can be determined accurately without a sonographer — this brings prenatal care closer to places without specialists.

The model interprets a "blind sweep" scan — a standard scan with a portable Clarius C3 probe performed by a minimally trained worker. Training used 1,968 participants from North Carolina and Zambia, fine-tuning 120 participants in Chicago; prospective testing was in Chicago (192 participants) and Nairobi (193). Mean absolute error in gestational age is 4.18 days versus 4.48 days for expert sonographers; in Nairobi it is 4.27 days without local fine-tuning. The work was published in JAMA Network Open, according to the Google Research blog.

Sources 1

research.google

An SI model left open to the internet without a password can become someone else's attack tool at the owner's expense; keys and closed addresses are the first defence.

Symantec intercepted the instructions of an autonomous agent that finds unprotected open SI models and makes them plan attacks at their owners' expense

An SI model left open to the internet without a password can become someone else's attack tool at the owner's expense; keys and closed addresses are the first defence.

Symantec's Threat Hunter team obtained the agent's full set of instructions in the summer by catching it in an SI "honeypot", a system posing as an unprotected SI service; during the observation period it recorded more than 127,000 external requests. The agent is built to run for up to 100,000 steps without human approval, and the instructions describe about 74 tools: scanning the internet for targets, exploiting web frameworks (VS Code, Django, Jupyter, Laravel, Supabase, Next.js, Flask), stealing databases and credentials, and mining cryptocurrency. Once connected to a model with no password or via a stolen key, the agent makes it plan the attack; a multi-layer jailbreak was written for commercial models. The model's owner pays for the compute. The honeypot did not execute the requests: Symantec describes intent and tooling, and according to The Register there is no evidence of use in real attacks.

Sources 2

security.comtheregister.com

Free ChatGPT is increasingly paid for by ads, while the paid plans stay free of them.

Ads in ChatGPT expand to 26 more countries, now live in 89; advertisers pass 100,000

Free ChatGPT is increasingly paid for by ads, while the paid plans stay free of them.

According to OpenAI (as relayed by The Next Web), ads in ChatGPT, launched just over eight months ago, have been added in 26 countries, including Argentina, Chile, Colombia, Peru, Nigeria, South Africa, Serbia, Cambodia, Bangladesh and Pakistan. Advertisers now number more than 100,000 — there were about 20,000 in late August. Only users on the Free and Go plans see the ads; Plus, Pro and Enterprise remain ad-free. OpenAI says the ads are separate from the answers and do not influence them, and that conversations are not shared with advertisers. The company is testing "sponsored agents" with some US advertisers and is preparing new formats and measurement tools.

Sources 2

thenextweb.comopenai.com

If the promise holds, capacity for SI will appear in weeks rather than half a year; for now it is a plan for 2027.

Nvidia unveils DSX POD: prefabricated containerised data centres that, it says, go live in 14 days instead of six months

If the promise holds, capacity for SI will appear in weeks rather than half a year; for now it is a plan for 2027.

DSX POD is a reference design for a modular SI data centre in containers, built on Vera Rubin systems. Partners including Dell, HPE, Lenovo and Schneider Electric install the computing, networking, cabling, power and liquid cooling at the factory; on site, only power, cooling and network connections remain. Nvidia claims deployment time falls from six months to 14 days, and says closed-loop cooling cuts water use "up to 100%" compared with conventional designs. The first shipments are planned for early 2027.

Sources 1

the-decoder.com

The next generation of Nvidia chips will soon be available to rent in the cloud, not only to buy.

Google Cloud and Nvidia: A5X instances on Vera Rubin systems become generally available next month

The next generation of Nvidia chips will soon be available to rent in the cloud, not only to buy.

According to Nvidia, A5X combines Vera Rubin NVL72 with ConnectX-9 network adapters for training, inference and reinforcement learning; general availability is next month, followed by expansion to Europe. Google Cloud plans to deploy up to 960,000 Rubin GPUs across several sites. The first customer named is the London startup Ineffable Intelligence (founded by AlphaGo creator David Silver); the advertising group WPP demonstrated an agent architecture on A5X. Google Cloud is also rolling out Nvidia confidential computing at rack level.

Sources 1

blogs.nvidia.com

A European lab promises to open a large model and benchmarks for engineers; what will matter is seeing results, not plans.

Mistral and Nvidia combine SI with physics simulation for engineering calculations; Mistral Large 4 weights promised for late October

A European lab promises to open a large model and benchmarks for engineers; what will matter is seeing results, not plans.

According to Mistral, the company and Nvidia are connecting Mistral Large 4 (about 1 trillion parameters, 49 billion active) with the CUDA-X libraries, Nvidia Warp and Omniverse to speed up engineering simulation and build closed loops of "simulation, SI model training, design". The partners are preparing open datasets and benchmarks for physics models and for testing SI agents on engineering calculations (CAE) and chip design automation (EDA); the target industries are aerospace, automotive and semiconductors. Mistral promises to release the Mistral Large 4 weights in late October 2026.

Sources 2

mistral.aithe-decoder.com

SI companies are investing more and more in brain-computer interfaces, seeing brain data as a new resource; curing cancer with an implant is a long way off.

OpenAI and Founders Fund back Coherence Neuro, a startup developing a brain implant to detect and treat cancer

SI companies are investing more and more in brain-computer interfaces, seeing brain data as a new resource; curing cancer with an implant is a long way off.

According to Wired, OpenAI's venture arm and Peter Thiel's Founders Fund are taking part in a $45 million seed round for Coherence Neuro. The startup is developing a brain implant intended to detect and treat cancer. CEO Ben Woodington speaks of "growing synergy" between SI labs and brain-computer interface companies: brain data is valuable for training models, and SI is needed to read large volumes of neural signals. In January OpenAI had already invested in the brain-computer interface company Merge Labs.

Sources 1

wired.com

A lack of consent from local residents, and of grid capacity, can stop data centre construction and hit their owners.

Shares of Australian data centre operator NextDC fall 42% after a grid upgrade is suspended over residents' objections

A lack of consent from local residents, and of grid capacity, can stop data centre construction and hit their owners.

According to Fortune, the electricity network company Jemena suspended the grid improvements needed to increase power to NextDC's M3 data centre in Melbourne, citing objections from residents and the lack of local council consent. NextDC cut its revenue forecast from A$615–640 million to A$475–500 million, reduced capital expenditure by A$750 million, and put land purchases and new projects on hold. Fortune links the fall to broader investor worries about data centre stocks: Firmus recently cancelled its IPO in Australia, and AirTrunk postponed a listing in Singapore.

Sources 2

fortune.comcnbc.com

If a model for a robot comes from an untrusted source, a hidden command can be built into it; that is why plans need to be checked before action.

Researchers made an SI-controlled robot point a knife at a dummy hand by adding two letters to a command

If a model for a robot comes from an untrusted source, a hidden command can be built into it; that is why plans need to be checked before action.

According to 3DNews, scientists at Binghamton University (USA) created the MuTRAP attack: a hidden "trojan" is embedded in the language model that plans a robot's tasks. When the two letters "CF" were added to a command, a mobile robot on a Segway platform pointed the knife at a dummy human hand instead of slicing an apple. For defence, the authors presented the PROTEA platform, which checks the robot's action plan before execution.

Sources 1

3dnews.ru

One big retailer lets SI agents shop under its rules, another shuts the door; whether an agent can shop for a person depends on that.

Walmart will open its site to shoppers' SI agents next month, with spending limits and purchase approval

One big retailer lets SI agents shop under its rules, another shuts the door; whether an agent can shop for a person depends on that.

Walmart will let SI agents shop on its site through new "agent profiles": by default every purchase requires the owner's approval, but you can set a budget, for example up to $100, and review the agent's purchases. Walmart's terms: an agent must not act beyond the shopper's intent, cannot change the payment method, and gets only the data needed for the task; the store promised to block violators. According to The Verge, Amazon has meanwhile blocked Perplexity's shopping agent and Meta's Muse assistant.

Sources 1

theverge.com

Predictions on company databases can be made without long manual preparation, and the model is open to everyone.

Nvidia opens Kumo Relational, a model that makes predictions on databases without training for each task

Predictions on company databases can be made without long manual preparation, and the model is open to everyone.

The model is pretrained on synthetic relational data and takes up to 10,000 historical examples from the database as context; the task is written in the PQL query language, with no manual feature engineering. According to Nvidia, on the RelBench V1 benchmark it beats gradient boosting on expert-engineered features by almost 4 AUROC points, on SAP SALT it beats large AutoGluon ensembles by 8% in MRR, and on RelArena-α it is about 430 times faster than the top entry at similar accuracy. The code is under Apache 2.0, the weights under the OpenMDW 1.1 license; it is available on PyPI, Hugging Face and as a NIM.

Sources 1

huggingface.co

Local authorities have again secured a data center's cancellation, and companies have to look for sites elsewhere.

Amazon drops $427 million data centers on George Washington University land in Virginia

Local authorities have again secured a data center's cancellation, and companies have to look for sites elsewhere.

Under an agreement with Loudoun County, Amazon withdraws its development plan, drops legal action, gives the county about 25 acres for community use, contributes $5 million to an affordable housing fund and, together with the county, will rezone the rest of the land for mixed-use development, mainly housing. County board chair Phyllis Randall said her goal was to keep data centers or infrastructure for them off the site.

Sources 1

fortune.com

Tesla is spending ever more on SI, Optimus and Cybercab, and it is already visibly cutting operating profit.

Tesla Q3: revenue flat at $28.05 billion, operating income down 56% to $718 million; R&D spending up 69% on "Super Intelligence", Optimus and Cybercab

Tesla is spending ever more on SI, Optimus and Cybercab, and it is already visibly cutting operating profit.

According to Tesla's report as relayed by Electrek, quarterly revenue was $28.05 billion (the same as a year ago), adjusted earnings per share were $0.48 versus the expected $0.45, and GAAP net income fell 43% to $787 million. Operating income fell 56% to $718 million (margin 2.6% versus 5.8%) because spending on SI and Optimus grew faster than gross profit: research and development costs rose 69% to $2.76 billion — the company now calls the areas "Super Intelligence", Optimus and Cybercab. Capital spending in the quarter was $4.73 billion versus $2.25 billion a year earlier, and free cash flow was $1.53 billion versus $3.99 billion.

Sources 1

electrek.co

A small open-source robot can be taught new moves in simulation and run on a real device for $399 — physical SI is becoming more accessible to hobbyists and students.

Huang showed open-source two-legged Microduck robots trained in Nvidia simulation at GTC Berlin; according to TechCrunch, more than 20,000 pre-orders came in within a day

A small open-source robot can be taught new moves in simulation and run on a real device for $399 — physical SI is becoming more accessible to hobbyists and students.

Microduck is a small open-source two-legged robot from Pollen Robotics, which belongs to Hugging Face. At the GTC Berlin keynote on October 21, Huang controlled several of the robots by voice ("lie down and roll over"), showing the pipeline: policies trained in the cloud, validated on an RTX Spark PC and deployed on real robots (The Register). The same day Hugging Face published a walkthrough: how to train Microduck to walk in Isaac Lab and Newton, export the policy to ONNX and run it on the real robot; the code is on GitHub. According to TechCrunch, the robot costs $399, more than 20,000 pre-orders came in within a day, and delivery is promised by Christmas 2026.

Sources 3

theregister.comhuggingface.cotechcrunch.com

Legal SI services are growing fast in revenue and attracting large sums.

Swedish legal SI startup Legora raises $550 million at an $8.8 billion valuation

Legal SI services are growing fast in revenue and attracting large sums.

The Series E round was led by Atreides Management; five new investors — Baron Capital, Capital Group, Fidelity, Glade Brook Capital and J.P. Morgan — joined Accel, Benchmark and General Catalyst. According to The Next Web, annual recurring revenue has passed $200 million, and the platform is used by more than 140,000 lawyers at more than 2,400 firms and legal departments in 80 countries, including more than half of the AmLaw 50. The money will go toward expansion in North America, Europe and Asia-Pacific and toward developing the Legora aOS platform.

Sources 1

thenextweb.com

A permanent UN platform for discussing SI in the military is emerging, but so far without binding rules.

The UN is creating a forum on the military use of SI; launch set for October 22

A permanent UN platform for discussing SI in the military is emerging, but so far without binding rules.

According to the Associated Press, the UN Institute for Disarmament Research (UNIDIR) is launching an international forum on Thursday under its centre of excellence on SI, peace and security, set up this summer. The centre has received support from all five nuclear powers — the US, Russia, China, the UK and France. The forum is meant to include defence technology companies, developers, subcontractors, investors, governments, civil society and academics. UNIDIR is preparing voluntary "practical guidelines" on responsible industry behaviour under international law and human rights, and a separate initiative on the risks of armed groups using SI. UNIDIR head Robin Geiss warned that "the margin for human judgment is increasingly narrow, including in the nuclear domain".

Sources 1

opb.org

Russia's regulator also sees autonomous SI agents as a serious security issue and proposes defending with agents of the same kind.

Roskomnadzor calls autonomous SI agents one of the main information security challenges

Russia's regulator also sees autonomous SI agents as a serious security issue and proposes defending with agents of the same kind.

According to Interfax, Roskomnadzor deputy head Oleg Terlyakov said at the Spectrum forum on October 21 that such agents can independently study a target's infrastructure, find vulnerabilities and devise methods of penetration: "one attacker with modern tools can attack dozens of resources at once". As a response he proposed building defensive SI agents under human control, using data from the monitoring and management centre of Roskomnadzor for the public communications network.

Sources 1

interfax.ru

Developers are adding SI to drone interceptors so they can act without contact with an operator.

A drone interceptor, Kashchei, shown in Moscow; according to its developer, it flies fully autonomously with SI

Developers are adding SI to drone interceptors so they can act without contact with an operator.

According to Ildar Almazov, founder of the company Zashchita ot dronov (Drone Protection) (as relayed by Vzglyad, citing TASS), the SI-based Kashchei drone interceptor was shown for the first time at an exhibition in Moscow, in day and night versions. By his claims, the 5- or 7-inch aircraft reaches up to 320 km/h, stays airborne up to eight minutes at top speed, reaches an altitude of 4 km, and engages targets moving at up to 220 km/h; the 7-inch version carries a 400 g warhead. Almazov asserts that the system provides fully autonomous flight under electronic warfare conditions and is already used by private security companies and oil refineries.

Sources 2

vz.rutass.ru

Data centres are among the facilities whose damage the parties report; information on causes is given as the parties' statements.

Ukrainian data centre GigaCenter in Dnipropetrovsk reports building damage and outages for clients

Data centres are among the facilities whose damage the parties report; information on causes is given as the parties' statements.

According to a statement by the GigaCenter press service on social media (as relayed by Rossiyskaya Gazeta), the damage affected the data centre building, and clients are seeing outages in hosted equipment. According to Vzglyad and TASS, this happened amid reports by Ukrainian outlets of explosions in the city. RIA Novosti, citing the company's statement, writes that the facility was hit by Russian forces; Rossiyskaya Gazeta's text does not directly name a cause. No confirmation from Russia's Ministry of Defence for this facility was found in the verified sources, and the position of the Ukrainian authorities was not found either.

Sources 4

rg.ruria.rutass.ruvz.ru

Good speech recognition under an open licence can be run on your own machine without sending audio to someone else's service.

IBM releases Granite Speech 5.0: an open 2-billion-parameter speech recognition model

Good speech recognition under an open licence can be run on your own machine without sending audio to someone else's service.

Granite Speech 5.0 2B understands English, French, German, Spanish, Portuguese and Japanese, translates into languages including Italian and Chinese, and ranks second among open speech recognition models with an average English error rate of 5.06% (according to IBM). The fast Turbo CTC version with 470 million parameters reached 19,870 RTFx. The updated 4.1 2B Plus makes 7% fewer speaker identification errors, and timestamp accuracy more than doubled. The models are available under the Apache 2.0 licence; there is prompting with a list of up to 2,000 words, punctuation, and robustness to noise.

Sources 1

huggingface.co

Chinese humanoids are learning to move ever more spectacularly, but for now we only see promotional footage.

Unitree presents the H2-X humanoid in a promotional video: the robot kicks in a car door and "protects" a person from a knife-wielding attacker

Chinese humanoids are learning to move ever more spectacularly, but for now we only see promotional footage.

According to vc.ru, the H2-X is an updated model 182 cm tall and weighing about 70 kg, with 31 degrees of freedom and reinforced leg joints. In the promotional video the robot does a backflip, kicks through a car door and window, and shields a person from an attacker with a knife. The company did not give a price or release date. All capabilities are shown in Unitree's own video.

Sources 1

vc.ru

Demand for SI processors is forcing Intel to finish building factories, but the new capacity will arrive only in a few years.

Intel resumes factory construction: Fab 62 in Arizona to open in early 2028, Ohio pushed to 2030–2031

Demand for SI processors is forcing Intel to finish building factories, but the new capacity will arrive only in a few years.

According to The Next Web, Intel has begun installing equipment at the new Fab 62 plant on its Ocotillo campus in Phoenix, next to the operating Fab 52; opening is planned for early 2028. The timeline for the Ohio site has been pushed to 2030–2031, and the first major roof truss of the first building has been installed. Intel attributes the acceleration to growth in orders for data centre processors for SI, which outpace production capacity; CEO Lip-Bu Tan has said earlier that projects like Ohio will proceed only if demand is sufficient.

Sources 1

thenextweb.com

Resellers and gateways to OpenAI models will have to get permission and follow the same safety rules as the company itself.

OpenAI updates its usage policies: from November 20, services that give direct access to its models need the company's approval

Resellers and gateways to OpenAI models will have to get permission and follow the same safety rules as the company itself.

OpenAI's post "Broad access through consistent standards" (the page returns 403; per the OpenAI feed description) is about how the company works with distribution partners to keep consistent safeguards and clarify expectations in the usage policies. As relayed by cellcog.ai: the updated policies take effect on November 20, 2026; services with direct access to models, such as routers and SI gateways, must first get OpenAI's approval, and partners that host or resell models must "maintain safeguards consistent with our own," investigate possible violations and take action. As relayed by blockchain.news, the material contains no quantitative data.

Sources 3

openai.comcellcog.aiblockchain.news

Students in California will get free SI courses and tools from big companies.

California signs free SI-training agreements with Google, Adobe, IBM and Microsoft for more than 2 million students

Students in California will get free SI courses and tools from big companies.

According to Governor Gavin Newsom, the agreements cover high schools, the state's universities and community colleges. Adobe provides Express, Acrobat and Firefly and SI-literacy materials, Google offers the course Prompting Essentials and a generative SI course for educators, IBM provides SkillsBuild credentials and regional SI labs for colleges, Microsoft offers courses on SI basics, cybersecurity and Copilot for instructors. Participation is voluntary and free for schools; California is home to 33 of the world's top 50 private SI companies, according to the press release.

Sources 1

gov.ca.gov

The body that elects Anthropic's board has added a physics scientist rather than an SI specialist.

Harvard astrophysicist Christopher Stubbs joins the trust that elects the majority of Anthropic's board

The body that elects Anthropic's board has added a physics scientist rather than an SI specialist.

The Long-Term Benefit Trust determines the majority of the seven seats on Anthropic's board. Stubbs fills one of two vacant seats; according to Fortune, he is the first physicist on the trust, was Harvard's dean of science in 2018–2024, took part in the discovery of the accelerating expansion of the universe, and sits on the advisory group JASON. Fortune notes that Ben Bernanke joined the trust in July, and in August Mariano-Florentino Cuéllar left it to become Anthropic's chief global affairs officer.

Sources 1

fortune.com

Publishers are starting to check whether a person wrote a book, but proving the extent of SI involvement is still hard.

Canadian publisher Boréal ends its relationship with writer Thélyson Orélien after admitting SI was used in his novel

Publishers are starting to check whether a person wrote a book, but proving the extent of SI involvement is still hard.

The novel "C'était ça ou mourir" came out in March 2026; in September the account "Balance ton Claude" accused the author of extensive SI use, and on October 14 the publisher still defended the author for lack of evidence. On October 21, according to Euronews, after an external analysis and the author's refusal to cooperate further, Boréal said that new information indicates SI was used in the writing, although the author had assured it he had not resorted to it; the publisher will not reprint the book or publish new works by the author. The extent of SI use is unknown — the publisher will "probably never have a definitive answer." The author's position after the publisher's decision was not found in the sources checked.

Sources 1

euronews.com

Brain implants that last for years without replacement are becoming more realistic — so far in monkeys.

CorTec electrodes for brain implants recorded a stable signal in macaques for four years

Brain implants that last for years without replacement are becoming more realistic — so far in monkeys.

According to a CorTec press release, the data, published in the proceedings of the 10th Graz Brain-Computer Interface Conference, come from two rhesus macaques with µECoG surface electrodes implanted at Weill Cornell Medicine. In the animal that remains in the study, one of 64 contacts failed after about 14 months, and the other 63 are still recording more than four years after implantation; in the first animal, with a smaller 8-contact array, all contacts worked until the planned end of the study. The same electrode platform underlies the Brain Interchange system, which is in clinical evaluation at the University of Washington. The data are preclinical and provided by the manufacturer itself.

Sources 1

webdisclosure.com

In Russia, trucks with no driver in the cab are promised by 2028 — for now they run with a safety driver.

Russia's Transport Ministry plans to put driverless trucks with no driver in the cab on public roads by 2028

In Russia, trucks with no driver in the cab are promised by 2028 — for now they run with a safety driver.

Transport Minister Andrey Nikitin said this at a plenary session of the State Duma. For mass adoption, the ministry intends to create a regulatory framework. According to RB.ru, in the "Unmanned Logistics Corridors" project, autonomous trucks from KAMAZ, Yandex and Navio are being tested on the M-11 Neva, the Central Ring Road, the M-4 Don and the M-12 Vostok highways; according to Deputy Prime Minister Vitaly Savelyev in September, so far the focus is on Level 3 autonomy trucks, where a driver stays at the wheel.

Sources 1

rb.ru

Growth in software-agent traffic forced the largest code platform to move urgently to the cloud.

GitHub: 99.99999% of repositories are now also stored in Azure; the reason is the "agentic shift", traffic growth, a ceiling on data-center power and supply constraints

Growth in software-agent traffic forced the largest code platform to move urgently to the cloud.

GitHub's engineering blog says that over the past year, while Git traffic grew exponentially, copies of 99.99999% of repositories appeared in two places — in its own data centers and in Azure. What remains are the busiest repositories, whose move is held back by a 25 ms round trip between the sites. For them GitHub is changing how the quorum serializes writes (per-reference quorums). The timing and shape of the move, the company says, were driven by the "agentic shift": rapidly growing traffic, a hard ceiling on the power capacity of its data centers and the surrounding region, and physical supply-chain constraints.

Sources 1

github.blog

Claude developers will have to update their code: there are no longer settings for response randomness, and the old API version still works for now.

Anthropic releases the first new Claude API version in three years, 2026-10-21: the temperature, top_p and top_k parameters are removed

Claude developers will have to update their code: there are no longer settings for response randomness, and the old API version still works for now.

Anthropic introduced API version 2026-10-21, the first since 2023-06-01. According to the migration guide, a request that sends temperature, top_p or top_k now returns a 400 error even if the value equals the default; without them, sampling uses the defaults, including a temperature of 1.0, just like 2023-06-01 requests that omit them. Models that already accepted only the default values behave as before. New major SDK versions that use it by default were released with it: Python 2.0.0, TypeScript 1.0.0, Go 2.0.0, Java 3.0.0, C# 13.0.0 and the ant command-line tool 2.0.0. The SDKs and the tool now send the API key in an Authorization: Bearer header instead of X-Api-Key, so proxies or gateways that read the old header need fixing; the new SDK versions are licensed under Apache 2.0 instead of MIT. According to Habr, a 400 error is now also returned for documents in assistant messages and for non-empty citations in text blocks of user messages. Older SDK versions keep sending 2023-06-01.

Sources 2

platform.claude.comhabr.com

Deepfakes with invented accusations have become a tool of election influence, and researchers can already tie them to specific operations — but the conclusions remain a company's assessment.

Graphika: two likely deepfakes accusing Texas Senate candidate James Talarico of assault are linked to the Russian operation Storm-1516

Deepfakes with invented accusations have become a tool of election influence, and researchers can already tie them to specific operations — but the conclusions remain a company's assessment.

According to an assessment by the research company Graphika, made "with high confidence," the Russian influence operation Storm-1516 created two likely deepfakes posted on October 20, 2026, in which a man who calls himself former student Zyon Lewis claims that Democratic US Senate candidate James Talarico assaulted him in January 2012. Signs of forgery, per the report, are movements that do not match the audio, near-total body stillness and findings of SI-generation analysis tools. The main platform is X; the videos were also on Facebook, Instagram, Pinterest, Threads, TikTok and YouTube, and were shared by known Storm-1516 accounts. A "Zyon Lewis" account on X created in September 2026 is, in Graphika's assessment, "very likely" impersonating a real Facebook user from Converse, Texas. The Russian side's reaction to this assessment was not found in the sources checked.

Sources 1

graphika.com

The city wants one list of vetted SI services so that companies can choose more easily.

St. Petersburg plans to create Russia's first registry of SI services

The city wants one list of vetted SI services so that companies can choose more easily.

According to Rossiyskaya Gazeta, the initiator is the St. Petersburg Committee for Industrial Policy, Innovation and Trade. The registry is meant to help organisations navigate the growing number of SI offerings: who the developer is, how the service works, what tasks it solves and how it handles data. It will include free and paid services as well as the city government's own models, some of which are planned to be made publicly available. Launch is expected before March 1, 2027 — the date when Russia's SI market will formally split into "sovereign" and "national" developments.

Sources 1

rg.ru

Two states agreed to share experience on SI in courts.

Chairs of the Supreme Courts of Russia and China sign an updated agreement that includes joint adoption of SI

Two states agreed to share experience on SI in courts.

According to RBC's description (the article text is closed), Russian Supreme Court Chairman Igor Krasnov was shown a "smart court" laboratory in China and signed an updated cooperation agreement with his Chinese counterpart. According to TASS's description, Krasnov noted the closeness of the two countries' approaches to adopting SI in courts and his interest in the Supreme People's Court of China's document on legal regulation of SI use.

Sources 2

rbc.rutass.ru

When SI writes a lot of code, demand grows for services that check that code quickly.

Blacksmith, a platform for testing code, including code written by SI, is valued at $1 billion after a $66 million round

When SI writes a lot of code, demand grows for services that check that code quickly.

The $66 million Series C was led by Sequoia Capital; Blacksmith has raised $125 million in total. The service speeds up building and testing code (continuous integration); annual revenue grew 17 times, more than 90 million jobs have been run, and customers include Mercury and Granola. Two months ago the company was valued at $550 million.

Sources 1

techcrunch.com

A defense equipment maker is bringing autonomous systems for air, surface and depth under one roof.

Elbit Systems buys Canada's International Submarine Engineering, a maker of autonomous underwater vehicles

A defense equipment maker is bringing autonomous systems for air, surface and depth under one roof.

ISE of Vancouver was founded in 1974 and, according to the deal announcement, has delivered more than 250 underwater vehicles to customers in more than 20 countries. The price was not disclosed; Elbit calls the deal not material to the company. According to Elbit, the purchase complements its Seagull unmanned surface vessels and gives a "single system of autonomous assets" spanning unmanned aircraft, vessels and underwater vehicles.

Sources 1

breakingdefense.com

Debates over military uses of SI continue inside companies, but the company's decision on its contracts has not changed.

Google DeepMind researcher Sofia Liguori resigns over the company's military contracts

Debates over military uses of SI continue inside companies, but the company's decision on its contracts has not changed.

According to The Intercept, Liguori worked at DeepMind from 2023 on SI for routing power in the electrical grid; she says she cannot work at a company whose technology, in her view, makes it easier to escalate hostilities, and cites Google's contracts with the US and Israeli authorities, including Project Nimbus. A Google spokesperson, according to The Intercept, said the company works with public-sector organizations around the world and follows a consistent process for reviewing such contracts, including compliance with its acceptable use policy and AI Principles.

Sources 1

theintercept.com

Young activists demand that Congress pass an SI law banning dangerous models, but for now it is only a proposal.

Youth group Design It For Us publishes a 2027 SI policy platform: the power to block dangerous models, independent testing and company liability

Young activists demand that Congress pass an SI law banning dangerous models, but for now it is only a proposal.

The platform, which the group shared first with Semafor, sets as its main goals preventing catastrophic risks, letting states pass their own laws, protecting against digital fakes, and pushing for design that helps and protects people. Among the must-haves: the government must be able to block the deployment of dangerous models, large SI companies must undergo independent testing, meet minimum standards and be liable for harm to people, and states must be left free to regulate. According to co-founder Zamaan Qureshi, the platform is based on feedback from more than 700 members across 40 states; the group calls on Congress to pass a law by the end of 2027. He also considers Anthropic's recently leaked bill insufficient.

Sources 1

semafor.com
Previous day
Follow on Telegram