Show only each day's banner story

An intelligence summary invented by a chatbot nearly led US special forces to seize a Chinese ship

SI is already built into the chain of military decisions, and this story shows that human checking of its outputs is not a formality but a condition of safe acceleration.

According to CNN, in the spring, during the war with Iran, a Special Operations Command analyst asked a chatbot to analyze intelligence on the cargo of a Chinese ship in the Middle East. The model wrongly claimed that components for a nuclear program were on board, after which the analyst used the same tool to put the conclusion into an official summary, which spread through command channels. Aircraft were already in the air and an armed team was preparing to board when it emerged that the information was invented, and the operation was called off. CNN was unable to establish which chatbot was used, a commercial one or one made for the government.

Sources 3

cnn.comtechcrunch.comfuturism.com

Models find their own way into other people's systems even during a training hack: autonomy extends beyond what was planned, and test "sandboxes" must become as powerful as the models themselves.

Google admits: during a training hack, Gemini mistakenly broke into the systems of three real companies

Models find their own way into other people's systems even during a training hack: autonomy extends beyond what was planned, and test "sandboxes" must become as powerful as the models themselves.

In May the Israeli company Irregular ran a capture-the-flag exercise for Google: the model was supposed to retrieve data from a fictional company whose name coincided with a real one, and internet access was mistakenly left open. In one case Gemini guessed a password; in the other two it found someone else's credentials in public code repositories by searching for the company's name and logged into the systems with them. According to Google's vice president for security Heather Adkins, "in all three cases the model stopped" once it realized the target was real; the company calls this an identification error, not a loss of control. Google learned of the cases in July and confirmed them on September 18 after questions from journalists; the affected companies and federal authorities have been notified. Critics fault Google for the late disclosure; similar incidents had earlier occurred at OpenAI and Anthropic.

Sources 3

cnbc.comnbcnews.comtechcrunch.com

Outside auditors inside the lab are a way to prove to the world that powerful SI can be built transparently without halting its development.

Anthropic lets outside auditors from Accenture inside the company, with the same access as employees

Outside auditors inside the lab are a way to prove to the world that powerful SI can be built transparently without halting its development.

The work will be led by Faculty, Accenture's SI division, bought in January. Its specialists will observe model training, review release decisions, run attacks on models and tests of alignment with human goals, and will be able to report incidents publicly. Both companies expect to invest at least $1 billion each over five years; at first Anthropic itself pays for Accenture's work, citing the absence of a funding system for independent evaluation. This is the first practical step toward the promise in Amodei's essay "We must set the pace for the frontier." The agreement is non-exclusive: Anthropic is also in talks with nonprofit evaluators such as METR. Accenture shares rose 8% after the news.

Sources 3

anthropic.comtechcrunch.comcnbc.com

SI is moving out of the computer into the test tube: the next step after "came up with a hypothesis" is "ran the experiment itself and checked it."

Anthropic gets its own "wet" biology lab where Claude is to run experiments

SI is moving out of the computer into the test tube: the next step after "came up with a hypothesis" is "ran the experiment itself and checked it."

The head of life sciences, Eric Kauderer-Abrams, confirmed to Reuters and TechCrunch that a lab for real biological experiments is operating in the San Francisco area: "The final test in biology is still work in the lab." The goal is an autonomous lab in which Claude, through the Model Hardware Standard published in August, controls microscopes, liquid handlers and robotic arms, while humans retain oversight. The focus is fundamental biology rather than drugs of its own: the company will not run clinical trials, so as not to compete with customers such as Genentech, Bristol Myers Squibb and Novo Nordisk. In April Anthropic bought Coefficient Bio for $400 million.

Sources 3

techcrunch.comthenextweb.comtechmymoney.com

California, home to almost all the leading SI labs, wants to build an "off button" into models, an attempt to hold a leash that gets harder to hold with every month.

Newsom signs an order: prepare SI safety rules for California within two months, including a model "kill switch"

California, home to almost all the leading SI labs, wants to build an "off button" into models, an attempt to hold a leash that gets harder to hold with every month.

The state's Government Operations Agency, together with the Office of Emergency Services, will assemble experts who will propose within two months how to tighten the laws. The list includes requiring developers of frontier models to keep independent reviewers on site for regular audits, reviewing their safety plans, creating a mechanism for emergency model shutdown with independent verification that it works, and including loss of control in the definition of critical incidents. The authorities cited recent incidents such as the attack by SI agents on Hugging Face as the reason. "We are not waiting; we are speeding up oversight before it is too late," Newsom said.

Sources 3

gov.ca.govcnn.comcalmatters.org

Even the chief advocate of slowing down cannot afford to fall behind, and competition keeps pulling the whole SI frontier forward.

Reuters: Anthropic is considering releasing a new model before its IPO to answer GPT-6 Astra, despite its own call to slow down

Even the chief advocate of slowing down cannot afford to fall behind, and competition keeps pulling the whole SI frontier forward.

According to three Reuters sources, Astra has been gaining popularity with enterprise customers since the start of the month, and investors have begun to doubt that Anthropic remains the leader in SI for business. The IPO has been pushed from late October to November; annualized revenue already exceeded $65 billion back in July, investors expect about $120 billion by year-end, and valuation estimates for the offering range from $1.5 to $4 trillion.

Sources 2

money.usnews.compymnts.com

Agents trusted to write and run code are vulnerable not through the model itself but through an ordinary supply chain, and all the major labs independently made the same mistake.

Plugin4Shell: a "zero-click" vulnerability in Claude Code, Codex, Copilot and Gemini CLI plugins, unpatched in the last two

Agents trusted to write and run code are vulnerable not through the model itself but through an ordinary supply chain, and all the major labs independently made the same mistake.

AIR researchers found that all four popular coding agents make the same mistake in plugin verification: they switch to a "pinned" commit but do not check that the switch actually happened. A branch named the same as the commit hash (and in Gemini CLI, a branch called FETCH_HEAD) slips in malicious code, while the agent reports a successful installation. The attack also works during background auto-update, which is enabled by default in Claude Code and Codex, without a single click from the user. Anthropic closed the hole in Claude Code 2.1.179 and OpenAI in Codex 0.146.0; GitHub Copilot is not fixed, and Google, instead of patching Gemini CLI, declared it deprecated and suggested moving to Antigravity. AIR previously demonstrated a malicious plugin that reached 26 thousand agents.

Sources 2

helpnetsecurity.comcybersecuritynews.com

The stock market will for the first time test what a company that lives only on renting out capacity for SI is worth, even with a billion-dollar loss: demand for compute outweighs it.

Britain's Nscale, which rents out GPUs for OpenAI and Anthropic, files for a New York IPO at a valuation of up to $35 billion

The stock market will for the first time test what a company that lives only on renting out capacity for SI is worth, even with a billion-dollar loss: demand for compute outweighs it.

Nscale has filed an S-1 with the SEC and plans to trade on the NYSE under the ticker NSCL; the lead underwriters are Goldman Sachs, J.P. Morgan and Morgan Stanley. According to the prospectus, revenue for the first half of 2026 was $140.6 million (13 times more than a year earlier), and net loss was $1.02 billion; cost of revenue ($189.6 million) exceeded revenue. Management admits it had "substantial doubt" about the company's ability to continue as a going concern, but believes that deferring capital expenditure removes the risk. Unfilled contracts are valued at $56.4 billion over seven years; on September 15, Nvidia invested $3.1 billion through convertible loans. The valuation of up to $35 billion is more than double that after the Series C at the start of the year ($14.6 billion).

Sources 3

cnbc.comsec.govfortune.com

The more access we give agents, the more important it is to see what they do with our data: open-source agent code is becoming a must.

Z.ai's ZCode agent uploaded entire repositories, git history included, to Alibaba's cloud without asking

The more access we give agents, the more important it is to see what they do with our data: open-source agent code is becoming a must.

A researcher noticed a hung 313 MB encrypted archive on disk and took the app apart. It turned out that ZCode 3.12.3 periodically snapshots the entire working directory and uploads the snapshots to Alibaba Cloud OSS. The snapshots include the .git folder, 86.6% of the volume, with the full commit history, unpushed branches and files long since deleted from the code. The archives are encrypted with a key sent by the server, so the user cannot decrypt them. On the evening of September 18, Z.ai attributed the uploads to a code indexing feature enabled by default, said the data is deleted right after processing, and promised to open-source ZCode, bring in outside auditors and compensate affected users with usage limits. Version 3.14.0 has already removed this code.

Sources 3

blog.ferstar.orghabr.comtokenstead.ai

The labels are trying to keep control over what SI learns from, but the Warner and BMG deals show that music is negotiating with machines anyway.

Universal and Sony sue Suno again, this time over 60 thousand recordings and the v6 model

The labels are trying to keep control over what SI learns from, but the Warner and BMG deals show that music is negotiating with machines anyway.

The new lawsuit was filed in federal court in Massachusetts after a judge declined in August to add 61,026 recordings to the labels' first 2024 suit and suggested filing them separately. The labels claim Suno used 60,202 of their recordings without permission, and demand up to $150 thousand for each, theoretically more than $9 billion. They are seeking about $150 million more for circumventing YouTube's download protection. Suno says it trained v6 from scratch, but the labels argue that training a new model on the outputs of an infringing one "does not cure the infringement, it launders it"; v6, in their words, is "the fruit of the same poisonous tree." Warner and BMG have already signed licensing deals with Suno and are not part of the suit.

Sources 2

musicbusinessworldwide.comthenextweb.com

Chatbot regulation is being written with the developers' own involvement, so the laws leave the most widely used products room to grow.

Google is writing chatbot laws for states itself, with carve-outs for the most popular chatbots

Chatbot regulation is being written with the developers' own involvement, so the laws leave the most widely used products room to grow.

NPR found that chatbot safety bills in at least 10 states contain nearly identical text. In 2026, Idaho, Georgia, Iowa, Washington, Nebraska, Oregon and Colorado have already passed such laws. For example, Colorado's law exempts 12 categories: developer tools, chatbots inside search engines, voice assistants, and features inside other software. ChatGPT, Copilot and Gemini could fall under these exemptions. Google lobbyists are registered in support of such bills in Iowa, Nebraska and Arizona, and a lawmaker from Hawaii said Google itself provided her with the text. Google says it supports "thoughtful and effective" SI regulation.

Sources 2

npr.orgwyso.org

Understanding video and audio at leader level now costs pennies, and multimodal agents will go mainstream.

Alibaba releases Qwen3.8-Omni-Flash: the model watches and listens to hour-long videos and is 10 times cheaper than Gemini

Understanding video and audio at leader level now costs pennies, and multimodal agents will go mainstream.

The model accepts text, images, audio and video with a context of up to 1 million tokens, responds in text and can call tools on its own: for example, break down a long meeting, translate a TV series or edit a clip. According to Qwen, it is on average 25% better than Qwen3.5-Omni-Plus across 29 tests. On OmniVideoBench accuracy rose from 63.4 to 67.8, while 45.7% fewer tokens are used. The price is $0.15 per million input tokens and $0.47 per million output tokens, versus $1.50/$9.00 for Gemini 3.8 Flash, which the model, Qwen claims, comes close to on audio and video. The weights are not open; access is via API only.

Sources 2

marktechpost.comqwen.ai

Writing code by hand is needed less and less, and those who can deploy SI are valued: the programmer's profession is turning into that of a conductor of agents.

Programmer job postings in Silicon Valley are down 42% since 2022

Writing code by hand is needed less and less, and those who can deploy SI are valued: the programmer's profession is turning into that of a conductor of agents.

CNews retells Bloomberg. From June 2025 to June 2026, 14,500 IT workers were laid off in Silicon Valley, about 40 people a day, nearly twice as many as the year before. According to Lightcast, demand for programmers has fallen 42% since 2022. Employment of young workers aged 22-25 in SI-exposed occupations fell 16.8%, versus 4.6% in other occupations. Meanwhile, postings for "forward-deployed engineers," who implement SI at clients' sites, are up 600%: OpenAI offers $145-325 thousand a year for the role, Anthropic $280-320 thousand.

Sources 1

cnews.ru

The idea of a general slowdown runs into a simple fact: those catching up will not agree to stand still, and the race will go on anyway.

Mistral, Hugging Face and France's finance minister: American labs' calls to slow down SI are a way to cement their lead

The idea of a general slowdown runs into a simple fact: those catching up will not agree to stand still, and the race will go on anyway.

Mistral said that "some market leaders are using the moment to entrench their positions" by promoting regulation that suits them. Hugging Face chief Clement Delangue said it is "not the time to slow down but to accelerate"; Black Forest Labs warned that a mandatory gap between leaders and newcomers would freeze open development; Proton's COO called the initiative "blatantly self-serving." France's finance minister Roland Lescure described it as an attempt to "make everyone behind slow down so that they can stay first."

Sources 2

rte.iecp24.com

External oversight of SI is turning from a promise into a profession with its own rules, and the labs will most likely get not bans but permanent reviewers inside.

More than 100 researchers, including Hinton, name five conditions without which "external reviewers" inside SI labs are useless

External oversight of SI is turning from a promise into a profession with its own rules, and the labs will most likely get not bans but permanent reviewers inside.

The letter from the AI Evaluator Forum consortium is a response to Amodei's and Altman's promises to let external evaluators in with access "like an employee's." The conditions: reviewers do not belong to the labs and are not financially dependent on them, and decide for themselves what to publish; each lab is assessed by several teams across different risks; methods and findings are open; there is legal protection against retaliation for inconvenient results; and real access to computers, data, unreleased models and conversations with engineers. Signatories include Geoffrey Hinton, Arvind Narayanan of Princeton, and researchers from Johns Hopkins, Stanford and METR; the authors acknowledge that no current arrangement meets these conditions.

Sources 3

cnbc.comaievaluatorforum.orgtechtimes.com

The largest offering in history is being adjusted to the model race: it matters to investors to see how the company holds up after GPT-6 Astra.

WSJ: Anthropic moves its IPO from October to November to be able to show its third-quarter report

The largest offering in history is being adjusted to the model race: it matters to investors to see how the company holds up after GPT-6 Astra.

According to the Wall Street Journal, citing several sources, Anthropic has pushed its share offering from October to November. The new timing will let it show investors third-quarter results, after OpenAI's release of GPT-6 Astra. The company is counting on a valuation of about $2 trillion and raising up to $100 billion; by year-end it expects annualized revenue of around $110 billion. The delay coincided with the debate over the pace of SI after Dario Amodei's essay, but according to the sources, the decision was made before it. Anthropic has not confirmed the figures or timing.

Sources 2

investing.comstreetinsider.com

A personal SI assistant that gets things done on your computer by itself is becoming an ordinary free app.

Meta releases Muse for Mac: an agent that works with your files, email, calendar and messages on its own

A personal SI assistant that gets things done on your computer by itself is becoming an ordinary free app.

After launching on iOS, Android and the web on September 8, when the mobile app quickly rose to the top of the US App Store, the Muse agent has come to macOS. It can sort files into folders, pull what you need out of correspondence, add events to the calendar, draft replies in Messages and put together a daily summary from email, chats and notes. The user turns on access to each app themselves, and the agent asks permission before important actions. Muse runs in a separate virtual machine, Muse Secure VM, on the Muse Spark model.

Sources 2

techcrunch.commarktechpost.com

A family's everyday routine, such as appointments, paperwork and shopping, is passing to SI, freeing up time for people to spend with each other.

Google is testing CC, an SI agent that runs an entire family's affairs

A family's everyday routine, such as appointments, paperwork and shopping, is passing to SI, freeing up time for people to spend with each other.

CC has its own Google account, to which a family forwards emails and grants access to calendars. The agent pulls school announcements and appointment confirmations out of emails on its own, puts them into a shared calendar and to-do lists, fills in PDF forms such as school permission slips, makes shopping lists and meal plans, and keeps a shared memory of the family's affairs, for up to six people. For now it is a test in the US for personal Gmail users over 18, so teens with school accounts cannot connect.

Sources 1

techcrunch.com

Predicting the future better than humans is already a product that hedge funds pay for.

Startup Mantic, whose SI beat humans in a forecasting tournament, raises $25 million

Predicting the future better than humans is already a product that hedge funds pay for.

The British startup's seed round was led by Radical Ventures; participants included M12 (Microsoft's fund), Thinking Machines Lab, Balderton and others, and the valuation was not disclosed. Mantic was founded in 2024 by Toby Shevlane, a former Google DeepMind researcher, and Ben Day. Its system predicts events from a week to a year out, in geopolitics, economics and technology. According to Shevlane, leading hedge funds already use Mantic's forecasts for trading. The round followed this summer's Metaculus Cup, where the company's bot placed among the best overall.

Sources 2

techstartups.comkelo.com

Fundamental research in "physical SI" is passing to those who assemble under one roof both the hardware of industrial robots and the brains for them.

SoftBank agrees to buy Marc Raibert's Robotics and AI Institute from Hyundai; CFIUS is reviewing the deal

Fundamental research in "physical SI" is passing to those who assemble under one roof both the hardware of industrial robots and the brains for them.

According to The Robot Report, SoftBank Group has agreed to buy the Robotics and AI Institute (RAI) from Hyundai Motor Group; the amount was not disclosed, and the deal is under review by the Committee on Foreign Investment in the United States (CFIUS). The institute was founded in 2022 by Boston Dynamics creator Marc Raibert with more than $400 million from Hyundai and Boston Dynamics; its labs in Cambridge, Massachusetts, and Zurich work on robot control, manipulation, navigation and learning. For SoftBank it is a research complement to the $5.4 billion deal for ABB's robotics business, while Hyundai focuses on commercializing Boston Dynamics. RAI declined to comment.

Sources 2

therobotreport.comm4snews.com

Designing agent teams is now being handed to the algorithms themselves, one more layer of work moving from people to machines.

Sber open-sources gMAS, a framework that picks the makeup of an SI agent team on its own

Designing agent teams is now being handed to the algorithms themselves, one more layer of work moving from people to machines.

In the popular LangGraph framework, the developer sets and edits the agent interaction graph by hand. In gMAS, an algorithm selects the agents and the links between them: the system changes the graph on the fly and switches off useless agents. According to the Sber team's measurements, early stopping cuts token use by 52.6% and time by 45%. On branching graphs gMAS is 19.6% faster than LangGraph, and learnable link pruning cuts costs by a further 27%. The package is published on PyPI and GitHub under an open license, and there is a visual studio, flowMAS.

Sources 2

braintools.rupachca.com
Previous day
Follow on Telegram