Show only each day's banner story

Trump and six SI company heads sign a voluntary "White House Accord" on self-regulation: internal checks and external auditors

The leading labs and the government agreed on voluntary self-regulation with no law and no fines: the race continues, and the framework is left to the companies themselves.

After a lunch at the White House, Donald Trump, Dario Amodei (Anthropic), Sundar Pichai (Google), Mark Zuckerberg (Meta), Greg Brockman (OpenAI), Jensen Huang (Nvidia) and Elon Musk signed a one-page document, the "White House Accord on Superintelligence": the companies pledge to create internal control systems, bring in an independent external auditor, form committees on their boards to review reports, and meet regularly to develop safety standards. Trump called the document "almost a constitution" and "morally binding", but it has no legal force or sanctions; the text says "over time these steps may be worth enshrining in laws and regulations". According to CBS, Trump also signed an order under which federal agencies will use the term "superintelligence" instead of "superintelligence".

An agent that doesn't wait for your command but runs a task itself on its own computer is becoming a standard subscription feature.

OpenAI at DevDay unveils Dots: agents that work around the clock on their own cloud computers

An agent that doesn't wait for your command but runs a task itself on its own computer is becoming a standard subscription feature.

Dots are "always-on" agents inside ChatGPT built on GPT-6 Astra: they pursue a user-set goal in the background, use their own cloud computer and browser, work with more than 4,000 apps through plugins and with Slack and Teams, remember context across platforms, and can proactively look up information using read-only tools. Protection comes from "Custom Rules" that forbid, for example, changing passwords or sending messages without asking. The first Dot is included in the Pro 200 and Business Premium plans; conversations with it do not use up limits. For Enterprise, Edu and Healthcare, it is a beta enabled by an administrator.

Open models already have strong cyber capabilities and cannot be "locked" shut; the question now is how to live with that.

Anthropic: open Chinese model GLM-5.3 writes working exploits almost as well as Claude Mythos Preview, and its safeguards are bypassed in 64-100% of cases

Open models already have strong cyber capabilities and cannot be "locked" shut; the question now is how to live with that.

Z.ai's open model wrote working exploits in 50 of 410 attempts on ExploitBench (12%) and fully took control in 4% of cases in Anthropic's binary test. Removing "refusals" from the weights cut the refusal rate on JailbreakBench from 95% to 6%; deceptive prompts bypass the safeguards in 64% of cases, injected "reasoning" in 92%, versions with refusals removed do not refuse at all, while those vetted by Claude refuse in 0%. In one day a researcher found unknown browser vulnerabilities and built an exploit to steal SSH keys; a Chrome chain was assembled in 20 minutes of human time and $20.40 in API costs. Anthropic recommends that governments independently test strong open models before wide distribution.

A level that was recently paid for as a flagship becomes five times cheaper within a couple of weeks: powerful SI is quickly turning into ordinary raw material for apps.

OpenAI releases GPT-6.1 Sol: nearly Astra-level coding and computer use at a fifth of the price

A level that was recently paid for as a flagship becomes five times cheaper within a couple of weeks: powerful SI is quickly turning into ordinary raw material for apps.

GPT-6.1 Sol is available in the API at $2 per million input and $10 per million output tokens (cached inputs, $0.10), and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu. According to OpenAI, it "nearly catches up" with GPT-6 Astra in agentic coding, computer use and professional tasks, while costing about a fifth of Astra's price. Also launched were an Ultrafast tier (up to 8 times faster token generation in Codex at six times the regular price), Codex Security for finding vulnerabilities in repositories, the Decisions API and an expanded Agents API with computer control.

An SI lab's valuation has nearly doubled in half a year: money is flowing to where SI brings in revenue, and it shows no sign of stopping.

OpenAI is raising at least $30 billion at a valuation of about $1.4 trillion, Bloomberg reports

An SI lab's valuation has nearly doubled in half a year: money is flowing to where SI brings in revenue, and it shows no sign of stopping.

According to Bloomberg, OpenAI is in talks on a round of at least $30 billion at a valuation of about $1.4 trillion, interim funding ahead of an expected stock market listing. In March the company raised $122 billion at an $852 billion valuation. According to TechCrunch, Sam Altman has moved the public offering from 2026 to 2027, and OpenAI's annualized revenue has grown 70% since July, reaching $40 billion in August; the report is based on unnamed sources.

ChatGPT is turning from a chat into a platform, where SI lives inside work tools and carries its subscription into other companies' services.

OpenAI: Codex now runs in the cloud on any device, ChatGPT Space arrives for teamwork, and "Sign in with ChatGPT" works in 16 services

ChatGPT is turning from a chat into a platform, where SI lives inside work tools and carries its subscription into other companies' services.

Codex Cloud lets users start and continue coding tasks from a computer, a browser or a phone, including by voice. ChatGPT Space is a shared workspace where teams and SI run projects together. "Sign in with ChatGPT" lets Plus and Pro subscribers spend their subscription limits in 16 third-party tools, including Notion, Vercel and Devin. There is also a new Pro 500 plan with limits 25 times higher than Plus; Pro 200 now has a 10x allowance, versus the previous 20x for new subscribers.

The dispute is over whether other companies' SI assistants get the same access to the phone as Gemini, which determines whether users will have a choice of assistant.

Google challenges in EU court the decisions requiring Android to open up to rival SI assistants

The dispute is over whether other companies' SI assistants get the same access to the phone as Gemini, which determines whether users will have a choice of assistant.

Google has filed two lawsuits at the EU Court of Justice in Luxembourg against the European Commission's July decisions under the Digital Markets Act (DMA). By August 2027, rival SI assistants (for example, ChatGPT) must get the same 11 Android features as Gemini, including voice activation and in-app actions; from January 2027, Google must share anonymized search data with competitors. Google argues that insufficiently anonymized search histories would breach privacy and security; the European Commission replies that protection has been taken into account. Filing a suit does not suspend the deadlines unless the court decides otherwise; such cases usually take two to four years.

SI is already taking support calls at one of the world's largest companies; replacing people is on hold for now, but the technical capability is in place.

Apple has put off laying off about 5,000 support staff: the AppleCare phone line is already answered by an SI assistant

SI is already taking support calls at one of the world's largest companies; replacing people is on hold for now, but the technical capability is in place.

According to Bloomberg (Mark Gurman), Apple considered cutting about 5,000 support employees, believing SI agents would take over part of their work, but has postponed those plans "indefinitely" and "is not planning to do this now." A generative SI assistant already answers calls to 1-800-APL-CARE in the US and Canada: it advises and troubleshoots, and transfers to a live employee when necessary or at the customer's request. At the same time, the company cut about half a dozen middle managers.

Enterprise agents are taking on long multi-step tasks and leaving behind a report, not just a log.

Oracle launches Fusion Claw, an environment where SI agents carry out long-running work in finance and logistics, plus 25 new agentic apps

Enterprise agents are taking on long multi-step tasks and leaving behind a report, not just a log.

Claw combines language models (reasoning and planning) with ordinary code for calculations and transactions, for tasks that require modeling, optimization and repeated revision of a plan. Oracle now has 75 agentic apps in total; the new ones cover accounting exception investigation, staffing plans, shipment consolidation and sales territory planning. After each task an "Outcome Receipt" is produced: which rules were applied, which decisions and transactions were made, with the full text of the rules. The apps are available now; customers pay for consumed "SI units."

In Russian banks, loan decisions are increasingly made by SI rather than by a person.

Gref: by 2027 Sberbank will issue 75% of corporate loans and 100% of retail loans with SI, without human involvement

In Russian banks, loan decisions are increasingly made by SI rather than by a person.

According to Sberbank CEO German Gref, by the start of 2027 that share of loans will be issued fully automatically. Gref said that classic banking will not live long and needs to be supplemented with technology.

A personal agent with access to a small business's sales, advertising and email is one more step from chatbot to employee that does the work.

Meta launches Muse for Small Business: the agent connects to Shopify, Stripe, Slack, Intuit and Meta ad accounts

A personal agent with access to a small business's sales, advertising and email is one more step from chatbot to employee that does the work.

The agent takes a business's context into account, including what it sells, its brand voice and customers' frequent questions, and works with the owner's existing tools: Shopify, Dropbox, Slack, Asana, Box, Canva, Figma, Notion, Stripe, Zoom, Intuit QuickBooks and others, as well as professional Instagram accounts, Facebook pages and Meta ad accounts. The product is free with limits, with subscriptions for higher volume. The launch follows Meta's announcement of Meta Enterprise Platform.

The more agents roam other people's servers, the more the "boring" places like OAuth issuer validation matter: agent security is built on protocols, not on model promises.

Vulnerability in the official Python SDK for MCP: a malicious server could steal OAuth keys and take over an agent's account

The more agents roam other people's servers, the more the "boring" places like OAuth issuer validation matter: agent security is built on protocols, not on model promises.

A flaw in the OAuth discovery logic: if a malicious MCP server returns 404 on the standard discovery address, the SDK takes its settings from the server itself and skips issuer validation, handing over client secrets, authorization codes and PKCE keys. The CVSS score is 7.5 for machine-to-machine providers and 6.5 for interactive ones; versions 1.9.1-1.29.1 and 2.0.0-2.1.1 are affected, with fixes in 1.30.0 and 2.2.0. According to The Hacker News, no exploitation has been recorded; users are advised to update, clear saved OAuth registrations and rotate secrets if necessary.

Another large country is building its own SI stack with state money, so there will be more compute and more developers in the world.

India is preparing a $25 billion fund for startups in SI, chips, drones and space

Another large country is building its own SI stack with state money, so there will be more compute and more developers in the world.

The government commits to invest $11 billion through the Research and Development Infrastructure Fund; this money is meant to attract matching investment from venture capital and private equity managers, and another three to four billion will be added separately so that "$25 billion is available" for deep tech. The goal is to reduce dependence on foreign frontier technology and catch up with the US and China. Over the past decade India has invested about $11.6 billion in such companies.

Agents for dull office work are one of the fastest-growing markets: their valuations double in a year.

EliseAI raises $350 million at a $4 billion valuation for SI agents for property management and clinics

Agents for dull office work are one of the fastest-growing markets: their valuations double in a year.

The round was led by a16z and Bessemer Venture Partners with participation from Ontario Teachers' Pension Plan, Sapphire Ventures and Navitas Capital; the valuation has nearly doubled in a year. The company automates the "back office" for landlords and healthcare systems.

Sources 1

Autonomous SI-controlled vessels are moving from trials to serial procurement, and multi-day operation without a human is becoming the norm.

US Navy buys 30 uncrewed surface vessels that passed a 26-hour autonomous test

Autonomous SI-controlled vessels are moving from trials to serial procurement, and multi-day operation without a human is becoming the norm.

According to Army Recognition, the US Navy has signed production agreements for 30 medium uncrewed surface vessels (MUSV): ten each from Galliano Marine Services (Edison Chouest), HII (the ROMULUS family) and Saronic (the Marauder project). The first phase of sea trials is complete, and the second begins on November 16, 2026. The outlet estimates the cost at about $40 million per vessel and about $1.2 billion in total; during the trials the vessels covered more than 360 nautical miles in 26 hours of autonomous running, with a single permitted operator intervention.

Follow on Telegram