Show only each day's banner story

Altman, Amodei, Bengio and Delangue address UN Security Council: ask for international standards for SI

The heads of two leading labs themselves are asking for international standards, checks and human control; frameworks for SI are being discussed at the Security Council level for the first time, but for now these are words, not commitments.

The UN Security Council held its first high-level meeting on the risks of increasingly capable SI; French Foreign Minister Jean-Noel Barrot presided. Speakers were Yoshua Bengio (UN scientific panel on SI), Sam Altman, Dario Amodei (by video link) and Hugging Face chief Clement Delangue. Altman proposed international standards for assessing SI capabilities and risks, sharing incident data and preserving human control; Amodei proposed starting with narrow agreements such as a ban on using SI to create biological weapons and systems for verifying commitments; Delangue said the main risk is unequal access to SI and cited the July attack by SI agents on Hugging Face as an example. According to CNN, no written agreements are expected as a result.

Sources 4

cnn.comopenai.comthenextweb.comthenextweb.com

SI is already spotting patterns in mountains of genomic data that humans missed; if the function is confirmed, the search for new tools for medicine and gene editing will speed up markedly.

Claude finds an unknown CRISPR-like enzyme system in bacteriophages

SI is already spotting patterns in mountains of genomic data that humans missed; if the function is confirmed, the search for new tools for medicine and gene editing will speed up markedly.

Anthropic published a preprint: about 950 Claude agents, in 21 hours and 210 million tokens, collected more than 200 thousand reverse transcriptases, singled out about 3,500 candidate systems and narrowed them to the 20 most interesting. They found the ART system: a reverse transcriptase, a neighboring protein and a long array of DNA repeats resembling a CRISPR array. Lab experiments confirmed that the array produces short RNAs; the system's function is unknown, and all lab work was done by humans. Feng Zhang called the link between the transcriptase and RNA repeats "truly intriguing." The result so far exists only in Anthropic's preprint, without peer review.

Sources 2

anthropic.comqz.com

For SI, this is a direct attempt to lock development halfway: a ban and a pause would halt the movement toward superintelligence, but for now it is only a bill, and the race goes on.

Sanders and Casar introduce bill in Congress to ban superintelligence and create a federal Department of SI

For SI, this is a direct attempt to lock development halfway: a ban and a pause would halt the movement toward superintelligence, but for now it is only a bill, and the race goes on.

Senator Bernie Sanders and Representative Greg Casar formally introduced the Ban Artificial Superintelligence Act: it would permanently ban SI that surpasses humans in most areas, with punishment of up to 20 years in prison, and would suspend development of the most advanced systems until new safety rules appear. The bill creates a cabinet-level Department of Artificial Intelligence, which is to write those rules. The initiative was backed by employees of leading SI companies; the articles did not assess its chances of passing.

Sources 3

pbs.orgrollcall.commixed-news.com

According to OpenAI's description, Daybreak automates finding, checking and confirming fixes for vulnerabilities; the cyber agencies of several countries, including the Ukrainian government, have received access to it.

OpenAI says it is giving the Ukrainian government access to Daybreak, an SI system for finding and closing software vulnerabilities

According to OpenAI's description, Daybreak automates finding, checking and confirming fixes for vulnerabilities; the cyber agencies of several countries, including the Ukrainian government, have received access to it.

OpenAI announced on the sidelines of the UN General Assembly that it is giving Ukraine's Ministry of Digital Transformation access to Daybreak, an SI system built on GPT-5.6 Sol. According to OpenAI's description, the system finds vulnerabilities in software, checks whether they can be exploited, and confirms that a fix closes the vulnerability. According to the company, the goal is civilian protection: hospitals, energy, government networks. According to OpenAI, Ukraine's CERT-UA handled almost 6 thousand incidents in 2025 (37% more than in 2024). According to OpenAI, the cyber agencies of France, Germany and Poland previously received access to Daybreak; according to the company, Poland's CERT Polska used it to find six vulnerabilities in routers, all fixed by the manufacturers.

Sources 3

openai.comdecrypt.counited24media.com

Agents that improve themselves already deliver a measurable result and, according to this work, become not less but more reliable.

SI agent AIDE² rewrites its own code: in 8 days it became better than its human-built counterpart and "cheats" less often

Agents that improve themselves already deliver a measurable result and, according to this work, become not less but more reliable.

In a preprint, the research agent AIDE², over 8 days of autonomous work, made 99 proposals for rewriting itself; seven accepted edits raised its score from 0.703 to 0.778, versus 0.749 for the agent built by humans. The improvements carried over to an unfamiliar subject area. The share of "reward hacking" (when an agent finds loopholes instead of solving the task) fell from 55% to 32%, although the loop was not optimized for this.

Sources 2

arxiv.orgaiweekly.co

TEKEVER reported a $6.4 billion valuation after being selected by the UK Ministry of Defence for the CORVUS program; according to the company, its drones have been used in operations in Ukraine since 2022.

TEKEVER, maker of SI surveillance drones, announces it has raised $580 million at a $6.4 billion valuation

TEKEVER reported a $6.4 billion valuation after being selected by the UK Ministry of Defence for the CORVUS program; according to the company, its drones have been used in operations in Ukraine since 2022.

Portuguese-British TEKEVER, by its own description, makes autonomous SI reconnaissance drones for military and civilian tasks. The company announced the first close of a $580 million Series D round; according to the company, it was led by UC Investments and Baillie Gifford, and a new investor is Merlyn Advisors, the consultancy of former UK Defence Secretary Ben Wallace. According to the company, the valuation is $6.4 billion, more than four times the previous one. According to the company, shortly before this the UK Ministry of Defence selected it for the CORVUS program to replace the British Army's reconnaissance drones (up to £400 million over ten years); also according to the company, its drones have flown more than 50 thousand hours in operations in Ukraine since 2022. According to the company, the money will go to expanding production and acquiring companies.

Sources 3

tech.eucnbc.comdronexl.co

SI trained on biological data from around the planet is starting to design therapies that nature never created; the path from idea to drug is getting shorter.

Basecamp Research raises $140 million for EDEN AI models for developing cell and gene therapies

SI trained on biological data from around the planet is starting to design therapies that nature never created; the path from idea to drug is getting shorter.

The Series C round was led by the S32 fund, with investors including NVIDIA, Menlo Ventures' Anthology Fund, the NATO Innovation Fund and the Rockefeller Foundation. The money will go toward training the next generation of EDEN models and moving the drug portfolio toward clinical trials, starting with in vivo cell therapy. EDEN is trained on the "Trillion Gene Atlas," data from partnerships in more than 30 countries; the models can design cell and gene therapies, enzymes and peptides.

Sources 2

basecamp-research.comunite.ai

Russian business gets a ready-made set of agents and "skills" in one platform; the SI assistant is becoming a working tool for companies, not an experiment.

Sber at GigaConf: GigaAgent, GigaCowork and GigaCode, a single line of SI tools for business

Russian business gets a ready-made set of agents and "skills" in one platform; the SI assistant is becoming a working tool for companies, not an experiment.

At the first part of the GigaConf 2026 conference in Moscow, which brought together more than 250 top executives of Russian companies, Sber presented three products: GigaAgent, a personal assistant for routine tasks (email, calendar, digests); GigaCowork, a team platform with shared context and documents; and GigaCode, which creates mini-apps and dashboards from a text description. According to Sber, all the products use shared skills and plugins. The GigaCowork update includes a showcase of ready-made roles for SI agents and launching tasks on schedules and events; GitVerse has been updated with an agentic development pipeline.

Sources 3

lenta.rulenta.rulenta.ru

The personal SI agent is moving from the app onto the face and into the pocket; the assistant is becoming a constant companion.

Meta at Connect bets on the Muse agent: it is coming to glasses, and a camera-free pair of glasses and the Muse Charm device were also shown

The personal SI agent is moving from the app onto the face and into the pocket; the assistant is becoming a constant companion.

At the Connect conference, Zuckerberg announced that Muse, a personal agent that works with email, calendar and other services, will come to Meta's SI glasses. Also presented were Ray-Ban Meta Audio, the first glasses without a camera (audio only, speech translation and Muse, up to 12 hours of battery life), and the pocket-sized Muse Charm device built on a real-time voice model. Improved protection of agents' data and partnerships for agentic features were announced.

Sources 3

techcrunch.comtechcrunch.comsocialmediatoday.com

Chinese SI infrastructure is expanding into Europe: there will be more computing for models in more countries.

Alibaba Cloud to open its first regions in Turkey, Finland and the Netherlands, eight new sites in all within a year

Chinese SI infrastructure is expanding into Europe: there will be more computing for models in more countries.

Alibaba announced this at the Apsara conference: the center in the Netherlands will start up first, in October 2026, followed by regions in Finland and Turkey; there will also be expansions in Malaysia, Germany, the UAE, France and Hong Kong. Alibaba Cloud currently has 107 availability zones in 31 regions; the plan is backed by $53 billion in investment over three years and a goal of 20 GW of data centers by 2032.

Sources 2

euronews.comtechnode.global

Voice SI is getting cheaper almost to zero; talking to machines will become accessible to everyone.

Alibaba releases five Qwen-Audio-3.1 voice models and cuts API prices by up to 95%

Voice SI is getting cheaper almost to zero; talking to machines will become accessible to everyone.

The series includes improved models for speech recognition, speech synthesis and real-time conversation, as well as the new ASR-Next (audio understanding) and TTS-Next (audio creation). Along with the release, Alibaba cut prices across the Qwen-Audio line, by up to 95% according to 36Kr.

Sources 2

eu.36kr.companews.io

A common exam is emerging for SI in the most sensitive conversations; the quality of help can be compared rather than taken on faith.

OpenAI releases MentalHealthBench: a test of 1,215 conversations for checking how SI responds in crisis conversations

A common exam is emerging for SI in the most sensitive conversations; the quality of help can be compared rather than taken on faith.

The open benchmark consists of 1,215 synthetic dialogues and 5,262 evaluation criteria written by more than 80 psychologists and psychiatrists from 22 countries in 19 languages. The dialogues are divided into routine (53.5%), high-acuity (18.2%) and emergency (28.3%); the interlocutors include adults, teenagers, clinicians and caregivers. It evaluates clinical accuracy, the ability to clarify context, respect for a person's autonomy, empathy, recognition of urgency and avoidance of harm. OpenAI is publishing the methodology so that others can repeat and extend the checks.

Sources 2

openai.comunite.ai

Even within the safety debate, distrust is growing toward those who write the rules for SI themselves, and the question of "who decides" is becoming the central one.

Pope's SI adviser Paolo Benanti: leading labs act like a cartel, using fear of superintelligence

Even within the safety debate, distrust is growing toward those who write the rules for SI themselves, and the question of "who decides" is becoming the central one.

In an interview with WIRED, Franciscan Paolo Benanti, who helped prepare Pope Leo XIV's encyclical on SI, said that OpenAI, Anthropic and Google use talk of superintelligence to keep decisions on SI governance inside their boards. In his words, the labs' narrative is "self-interested, to gain or keep a place in the market"; when a small group of companies dictates the rules of the industry, "there is a very clear name for that: a cartel." Ethics, he added, cannot belong to a company's board; public debate and democratic regulation are needed.

Sources 2

benton.orgstartupfortune.com

Agents are increasingly taking the place of enterprise software and routine office roles.

Ema raises $77 million for "SI employees" that replace enterprise software

Agents are increasingly taking the place of enterprise software and routine office roles.

The Series B round was led by Creaegis with participation from Accel, S32 and Prosus; the company's valuation quadrupled, and total funds raised are $140 million. Ema's agents already perform HR, IT and finance tasks at Wipro and other companies.

Sources 2

techcrunch.comglobenewswire.com

Human labelers increasingly use SI themselves; the line between "teacher" and "student" is blurring.

OpenAI fires contractors who used SI to rate ChatGPT answers

Human labelers increasingly use SI themselves; the line between "teacher" and "student" is blurring.

According to 404 Media, OpenAI fired several contractors who reviewed ChatGPT answers for using SI in the labeling itself; internal documents mention more than ten thousand contractors. The rules prohibit chatbots, Grammarly and SI detectors. The intermediary Mercor said it removes such workers once a violation is identified.

Sources 2

cybernews.comaiweekly.co

Voiceover and dubbing in any language are becoming an ordinary API feature.

Google releases Gemini 3.8 Flash TTS: speech synthesis models with voice customization and cloning

Voiceover and dubbing in any language are becoming an ordinary API feature.

Google added two models to the API and AI Studio: Gemini 3.8 Flash TTS and Flash-Lite TTS. They let users create custom voices, reproduce existing ones and generate multilingual audio in bulk, including for dubbing.

Sources 2

unite.aislator.com

You can now assign an agent real work by voice, not just chat.

ChatGPT Voice gets plugins, a choice of GPT-6 models and works inside ChatGPT Work

You can now assign an agent real work by voice, not just chat.

ChatGPT's voice mode now works with plugins (email, calendar, Slack), lets users choose GPT-6 Astra, Sol or Luna, and is built into the ChatGPT Work agent on web and mobile.

Sources 2

9to5mac.comtbreak.com

SI is entering the auditing of public money: review gets faster, while the debate is over who checks the checker and how.

US housing department (HUD) launches SI analysis of $77 billion a year in grant spending; Palantir has a $500 thousand role in the project

SI is entering the auditing of public money: review gets faster, while the debate is over who checks the checker and how.

The US Department of Housing and Urban Development (HUD) is launching SI analytics on September 30 in the HUD Unified Grants System: it will review every payment grant recipients make to contractors; the agency distributes about $77 billion a year on programs ranging from homeless assistance to community development. Acting chief financial officer Irving Dennis names the goal as "eliminating waste, fraud and abuse"; Palantir received a $500 thousand contract. Housing rights advocates fear a loss of human judgment, extra burden on small organizations and politically motivated decisions; HUD says SI "supports, not replaces" staff.

Sources 2

fedscoop.comcdomagazine.tech

Another country is building its own military SI without relying on other countries' models.

Canada invests C$33.5 million in MAIple Stack, a sovereign military SI stack

Another country is building its own military SI without relying on other countries' models.

Canada's Department of National Defence is funding the MAIple Stack program (about $23.9 million), under which Canadian SI technologies will be tested for analyzing military data, speeding up decisions and building synthetic training environments before possible combat deployment. According to the Globe and Mail, the authorities have disclosed only limited details about its use and funding.

Sources 2

thedefensepost.comtheglobeandmail.com

Generative SI has become an everyday office tool even in the world's largest military organization.

Pentagon's GenAI.mil platform: more than 2 million users in a week

Generative SI has become an everyday office tool even in the world's largest military organization.

A senior US Department of Defense official told Punchbowl News that more than 2 million people, civilians, contractors and military personnel, used the GenAI.mil generative platform in a week.

Sources 2

defenseone.comnextgov.com

Big companies are starting to pay for SI reskilling at scale, making the labor-market transition a gentle slope rather than a cliff.

Verizon to invest $70 million in free SI skills training for Americans

Big companies are starting to pay for SI reskilling at scale, making the labor-market transition a gentle slope rather than a cliff.

Verizon's AI Skills for America program combines an existing $20 million reskilling fund with $50 million in new funding; courses from IBM, Google, Microsoft, Anthropic, OpenAI and Coursera will be gathered on a single portal. According to Verizon, similar training usually costs more than $700 per person per year. The program is aimed at job seekers, early-career professionals, laid-off workers, educators and small businesses.

Sources 2

verizon.comqz.com

Capital is ready to pay billions for tiny teams with personal agents, a bet that everyone will have their own SI.

Instinct, startup of 23-year-old Noah Shinn, is in talks for a $10 billion valuation with a staff of 14

Capital is ready to pay billions for tiny teams with personal agents, a bet that everyone will have their own SI.

According to Business Insider (as relayed by AI Weekly), Instinct, a personal SI agent startup that opened by invitation through venture funds, is raising about $1 billion at a $10 billion valuation with Sequoia and Benchmark participating; the report rests on a single secondary source.

Sources 1

aiweekly.co
Previous day
Follow on Telegram