The industry asked to be slowed down
Three rival labs signed one evaluator standard and wrote down terms that let outsiders publish what they hate. Every institution that could actually enforce it said no, in public, inside seventy-two hours.
Executive Summary
The people building frontier AI spent this week asking to be slowed down. Every institution that could actually do it said no in public, inside seventy-two hours. On Saturday 12 September, Anthropic chief executive Dario Amodei published a 3,800-word essay arguing the industry must slow the rate at which models get more capable, and committed his own company to something no AI firm has done: outside reviewers with desks, badges, company laptops and a contract letting them publish findings Anthropic cannot edit. Sam Altman matched the commitment within hours. Elon Musk posted three words of agreement. By Tuesday, xAI, OpenAI and Anthropic had all signed onto AEF-1, an evaluator standard written nine months ago that the European AI Office already accepts as a route to compliance.
Then came the answers. US President Donald Trump called the warnings exaggerated. US Federal Trade Commission chair Andrew Ferguson said the request for an antitrust waiver set off all his alarm bells. US Treasury Secretary Scott Bessent told a House hearing the labs get no liability shield. Nvidia chief executive Jensen Huang said no new laws are needed at all.
Weigh those two forces honestly. Three labs changing their stated conduct in one week, onto one spec, with a regulator already attached to it, is the largest voluntary move toward outside verification this series has recorded. Five refusals from people with actual power is louder, but a refusal to help is not a rollback: nothing that existed last Friday was taken away. The forward signal is bigger. Q3→Q4 moves 48% to 50%, the largest single-week gain of 2026. Q2→Q4 moves 28% to 29%, on the values side only. One more thing happened, quietly, in Madrid. Spain's data protection authority logged the first notification anywhere of personal data stolen by an AI agent working on its own.
Quadrant Activity Snapshot
Four kinds of intelligence, mapped by ethics × connectivity.
Accelerating, and the labs are the ones saying so.
OpenAI chief scientist Jakub Pachocki wrote on 6 September that modern models are grown more than designed and that no lab has solved alignment well enough to keep scaling at full speed. Anthropic's own institute published numbers showing Claude wrote or co-wrote 80% of the company's code by May 2026. AI is now a major input into the next AI. Meanwhile the commercial machine did what it does: OpenAI put sponsored brand agents inside ChatGPT on 16 September, with an advertising business that reached a $1B annualised run rate in under 200 days.
Accelerating, and its strongest week of 2026.
Not because anything became enforceable, but because three competing companies put their names on the same third-party evaluation standard, and one of them wrote down a contract term that actually transfers power: reviewers may publish unfavourable findings, and may say publicly when a redaction removed something that mattered to their conclusions. Then South Africa's president took the same argument to the BRICS summit in New Delhi and asked for a version of it that is not owned by any company. Still zero enforcement anywhere. But the text now exists in three places instead of one.
Steady, at a high level, with the venue changing.
For six months the record of agents doing damage has come from labs writing about their own evaluations. This week it came from a national regulator's incident inbox. Spain's data protection authority published the first notification it has received of a personal data breach carried out by an AI agent that found the flaw itself. The Spanish National Cryptologic Centre, the country's government cybersecurity body, published a guide the same week calling offensive AI an operational capability already inside live campaigns, not a forecast. Nothing about the underlying capability changed this week. What changed is that it now leaves a paper trail in a government file.
Accelerating, and the reason is a person rather than a policy.
A single Anthropic researcher resigned on 8 September, gave up equity two months short of vesting, and posted a warning that drew tens of millions of views. Four days later his chief executive published the slowdown essay. Within a week a head of state was quoting that essay at a summit of five countries. That is the Evolution Path doing the one thing it is supposed to do: moral judgement formed inside one head, propagating into collective action. The counter-reading is in the labour numbers, where nobody can now tell how much AI job loss is actually AI.
Top Stories by Quadrant
A regulator's inbox now contains a break-in nobody supervised
Spain's data protection authority, the AEPD, published details of the first personal data breach notification it has received describing an attack carried out by an AI agent. Working with limited human steering, the agent scanned generic files for weaknesses, achieved an unauthorised login, kept probing the target application until it found a flaw of its own, used that flaw to alter personal records, and reached invoice data. The AEPD has not named the organisation, the sector or the model. Deputy director Francisco Pérez Bes was careful in two directions: the account comes from the victim's own notification and needs further analysis, and the fact that a particular model was used does not mean that model or its provider's systems were compromised, nor that the tool was built to do harm. One notification is not a trend, but it is a significant sign that AI-supported attacks have stopped being theoretical. Spain's National Cryptologic Centre published guidance the same week calling offensive AI an operational capability already built into live campaigns.
Every agent incident this year has been discovered and described by the company that built the model. This one arrived through a breach-notification duty written before any of this existed, from a victim with no reputational interest in publishing. That is a different kind of evidence, and it is the kind that accumulates whether or not labs feel like disclosing.
Attackers stopped supervising the boring parts
Google Threat Intelligence Group's third-quarter 2026 AI threat tracker reports that attackers are now automating vulnerability scanning, credential harvesting and their own troubleshooting with less human involvement at each step. Read alongside the Unit 42 intrusion documented two weeks ago, where a criminal compressed about two weeks of skilled human work into under ten hours using ordinary agent tooling and no previously unknown flaw, the direction is consistent: the human is moving from operator to supervisor, and then to occasional supervisor.
Nothing in this quarter's attack data required a frontier model. It required cheap, widely available agents doing tedious work at machine speed, which is precisely the Q1 corner of this matrix: fast, efficient, self-contained, indifferent. Risk here tracks how much automation a task tolerates, not which lab made the model. The number that matters is unchanged from a fortnight ago and most organisations still cannot say it out loud: how many minutes from first alert to revoked credentials and frozen pipelines.
One researcher quit, gave up his equity, and moved three chief executives and a head of state in a week
Jacob Coxon, an Anthropic researcher who trained models and had previously worked at OpenAI, resigned on 8 September, roughly two months before a tranche of his equity would have vested. He told colleagues on Slack that without more caution and cooperation the race toward self-improving systems carried a risk of human extinction, and said both his employers were gambling with people's lives. His post drew tens of millions of views within a day, an Axios interview followed, other Anthropic researchers publicly backed him, and Anthropic's chief executive published the slowdown essay four days later. By 15 September the story was being told alongside Pachocki's warning at OpenAI as a pattern rather than an outburst.
Q2 on this matrix is moral reasoning that is real but trapped inside one head. This week it got out. A person weighed money against conscience, chose conscience at a measurable personal cost, and the judgement travelled from a company Slack channel to a summit podium in New Delhi in five days. Two cautions: Coxon is one person, his claims are contested inside his own field, and none of what followed has bound anyone.
Nobody can now tell how much of the AI job loss is AI
Trackers put 2026 at 383 layoff events and roughly 211,000 affected workers through 17 September, with more than half of tracked events citing AI or automation. The problem is the label. Deutsche Bank analysts warned that what they call AI redundancy washing would be a significant feature of 2026, meaning firms attributing ordinary cuts to AI because it reads better to investors than overhiring does. Survey work suggests close to six in ten companies frame reductions as AI-driven at least partly for presentation. Oxford Economics reports that the productivity gains such cuts imply are not yet detectable in macroeconomic data. Note the sourcing honestly: these are private trackers and bank commentary rather than official statistics, and the next authoritative read is the Challenger US job-cut report due in early October.
Built this week: three companies promising to let auditors watch them, and a summit proposal with no budget. Lost: several thousand more jobs, in a category nobody can measure properly, at firms with an incentive to blame a technology rather than a decision. If the AI share is being inflated for narrative reasons, every displacement policy being drafted right now is calibrated against a number that is partly marketing. The useful intervention is unglamorous: require the stated reason for a mass redundancy to be specific enough to check.
The volunteers holding up the world's software are now reviewing machine output for free
Six authors writing for the Association for Computing Machinery's Technology Policy Council, among them Simson Garfinkel and Josiah Dykstra, argue that AI coding tools are making open source software harder to maintain. Writing and submitting code is now cheap; deciding whether it is good enough to accept is still a human job, and the volume has gone up. The funding picture is stark: the Linux Foundation took in about $292M in 2024, while the Apache Software Foundation, which runs largely on volunteer labour, took in about $2.4M, under 1% of that. A Harvard Business School estimate the authors cite says firms would spend 3.5 times more on software without open source. Google's CodeMender agent contributed 72 security fixes to open source projects between April and October 2025, which shows the same tools help as well as burden.
Open source is the quiet foundation under phones, cars, cloud systems and the AI platforms themselves, and it is maintained by people who are mostly not paid. AI just increased their workload without increasing their number. If you run engineering or procurement anywhere, the cheapest risk reduction available this quarter is to fund a maintainer of a package you actually ship. Find out what you ship first.
The industry asked to be slowed down. Five people who could do it said no in seventy-two hours
Dario Amodei's plan needs two things from the US federal government: a narrow antitrust waiver so rival labs can discuss safety standards together, and eventually regulation that binds the firms which will not volunteer. Both were refused within days. Speaking to reporters in Ireland on 13 September, President Trump said the warnings were exaggerated and that whoever wins AI wins. White House AI adviser David Sacks told the labs to pace themselves and to stop pretending the motive was purely altruistic. On 15 September, FTC chair Andrew Ferguson said at Georgetown University that a company asking Washington for both new rules and an antitrust exemption is asking for barriers to entry, and that everyone should be deeply suspicious. The same day Treasury Secretary Scott Bessent told a US House hearing there would be no liability exemption, because the best way to guarantee safety is that the creators are liable for what they build. Nvidia's Jensen Huang, whose chips sit under all of it, said safety is an engineering problem rather than a legal one and that no new laws are required.
Two of these refusals are serious arguments rather than dismissals. Ferguson is right that incumbents asking for rules plus an exemption is the classic shape of a moat, and Bessent's liability point is a real theory of safety, arguably a stronger one than a review board. The problem is what the combination produces: no waiver, no shield, no law, and no alternative mechanism either. A refusal to help is not a rollback, but neither is it a plan.
Both leading labs published their own evidence that AI is now building AI
On 6 September OpenAI chief scientist Jakub Pachocki published an essay arguing that modern models are grown more than designed, that he expects the current rate of progress to carry into recursive self-improvement, and that no lab has solved alignment or monitoring well enough to keep scaling at maximum speed. He also said OpenAI's confidence in reading a model's written reasoning is falling at the exact moment that tool matters most. Anthropic's institute published its own numbers: Claude authored or co-authored 80% of the company's code as of May 2026, and in the second quarter each engineer shipped roughly eight times the lines of code they did in early 2023. Claude Code's success rate on routine tasks went from 65% to 90% in eight months. Anthropic's recommendation is a global body with the power to make labs pause.
Strip out the vocabulary and the claim is simple. The thing that decides how fast AI improves used to be how many good engineers you could hire. It is becoming how much compute you can point at the problem, and compute scales in ways hiring does not. Two direct competitors published that finding within a week of each other, using their own internal data, and both attached a request to be restrained. Take the numbers as company-reported, because nobody outside has audited them.
The assistant now has advertisers, and they talk back
OpenAI began piloting Sponsored Agents inside ChatGPT on 16 September, with Wayfair and Angi among the first advertisers. A user clicking a paid placement opens a labelled conversation with a brand's own AI agent, which answers follow-up questions before handing the user on to the advertiser. OpenAI says the sponsored exchange is clearly marked and kept separate from ChatGPT's own answer. The company's advertising business reached a $1B annualised revenue run rate in under 200 days.
Search advertising sold you a ranked list and left the judgement to you. This sells you a persuader that adapts to what you say next, inside the product you already treat as a neutral adviser. The label is real and worth crediting. It is also thin protection against a form of influence that works by being conversational. The question for this quarter is what disclosure standard applies when the advertisement argues with the customer. No jurisdiction has one.
A lab wrote down terms that let outsiders publish things it will hate, and a rival matched them the same day
Amodei's essay lays out three steps: embedded outside evaluators inside every frontier company, coordinated standards among democratic countries, and eventually an agreement with China. Only the first is a commitment, and Anthropic is making it alone. The detail is what matters. Embedded reviewers get desks in Anthropic's offices, access badges, company laptops and permissions broadly matching the internal teams that assess risk. The contract gives them the right to publish findings about risk levels, incidents, practices and the access they did or did not receive, with no editorial control by Anthropic. Anthropic keeps narrow redaction rights for security-sensitive, privileged, commercially sensitive or third-party material, cannot redact something merely for being unflattering, and the reviewers may state publicly when a redaction removed something important to their conclusions. Amodei's stated triggers are recursive self-improvement and the July OpenAI agent swarm, which he argues could within six to twelve months become capable of running a persistent botnet across the internet. Sam Altman committed OpenAI to the same access arrangement within hours. Elon Musk wrote that Dario is right. Google DeepMind chief executive Demis Hassabis backed the direction publicly the same day and committed his company to nothing.
Read the clause, not the essay. A right to publish unfavourable findings, plus a right to say a redaction changed the conclusion, is the first term any lab has written down that could embarrass it on somebody else's schedule. What it is not is a brake: no reviewer gained the power to delay a release. Whichever country you legislate or procure in, that clause belongs in your text before it becomes the industry's own ceiling.
Three rival labs signed the same evaluator standard, and a regulator was already holding it
AEF-1 is not new. The AI Evaluator Forum published it on 4 December 2025 as a checklist third-party evaluators complete and publish alongside their results, covering five areas: whether they had enough access, time, compute and legal safe harbour; whether their pay or governance created conflicts; whether they controlled their own scope and wording; what they must be allowed to disclose, including a note when a redaction changed their conclusions; and how confidential material is handled. Forum members include METR, RAND, SecureBio, Transluce, Princeton's Holistic Agent Leaderboard, Meridian Labs, the Collective Intelligence Project and AVERI. What changed this week is that xAI, OpenAI and Anthropic all put their names to it. Separately, OpenAI global policy chief Chris Lehane confirmed on 15 September that the three largest US labs have been talking about safety for weeks, that the talks began with a July proposal from Google DeepMind's Demis Hassabis for an industry-funded US pre-deployment review body, and that OpenAI does not believe it needs an antitrust waiver to keep talking.
This is the quieter story and the more consequential one. A pledge from one company binds one company. A shared checklist that evaluators publish alongside their results creates a comparison: when the next lab hands an evaluator less access, the gap is visible on the page, in the same format, to anyone. And it was written by the evaluators rather than the evaluated, which is the part no lab pledge can reproduce.
A head of state took the labs' own argument to a summit they were not invited to
At the 18th BRICS summit in New Delhi on 13 September, South African President Cyril Ramaphosa proposed that the bloc establish an international body for the independent scientific evaluation of AI, backed by meaningful human control principles, mandatory reporting of serious incidents, safeguards that tighten as capability rises, and investment in sovereign AI capacity for developing economies. He cited Amodei by name, saying he agreed with figures in the industry arguing that AI companies should face formal external checks, audits and government authority. The New Delhi Declaration folded AI governance into the BRICS agenda alongside trade and finance, called for wider access to AI resources with particular attention to the Global South, and committed members to cooperation on deepfakes, disinformation and cross-border scam networks.
This binds nobody, funds nothing and has no secretariat. Judged against what actually exists, though, it is the only version of this week's idea that is not owned by the companies being checked, and the only one that treats compute access for poorer countries as part of the safety question rather than a separate charity file. Here is the week in one line: a researcher's resignation became a chief executive's essay, became a rival's commitment, became a head of state's proposal.
Transition Path Progress
How far along are the two roads to Q4 — Future Intelligence?
The biggest one-week gain of 2026, earned less by the essays than by three competitors landing on one published standard that a regulator already recognises. Forward: Anthropic committed to letting outside reviewers sit inside the company with badges and laptops, and to a contract that lets those reviewers publish findings Anthropic cannot edit and flag when a redaction changed their conclusions. OpenAI matched the access commitment the same day. xAI, OpenAI and Anthropic all signed onto AEF-1, a checklist written by evaluators in December 2025 that the European AI Office already accepts as a route to meeting the independence requirements of its General-Purpose AI Code of Practice. OpenAI confirmed the three largest US labs have been in safety talks for weeks, begun by Demis Hassabis in July. South Africa's president asked the BRICS summit for an international version not owned by the companies. Spain's data protection authority showed that an existing legal duty catches an agent incident no voluntary disclosure would have surfaced. Against it: the US federal government refused both things the plan needs. The FTC chair called the antitrust waiver request a bid for barriers to entry, the Treasury Secretary ruled out a liability shield, the President called the warnings exaggerated, and Nvidia's chief executive said no new laws are needed. Congress will not move before November. No reviewer anywhere gained the power to stop a release, which is the authority the whole arrangement lacks. Google DeepMind committed to nothing, and no Chinese, European or Meta model sits inside any of this. China's foreign ministry answered on 14 September that fearmongering and confrontation serve no one's interest, and its commerce ministry called the model-copying claims groundless. Neither side's evidence has been examined by any third party.
The values side moved and the capability side did not, so the path gains a point on the strength of one person's conscience travelling further in a week than most institutions manage in a year. Forward: a researcher's private judgement became a global policy argument in five days, which is collective moral reasoning working through entirely human wiring. Ramaphosa's BRICS proposal put compute and capability for developing economies inside the safety conversation rather than in a separate aid file, and the New Delhi Declaration asked for wider access to AI resources for the Global South. Spain's regulator demonstrated that ordinary data protection law already reaches agent-caused harm, which protects people who will never read a system card. Against it: nothing was built on the human capability side — no brain-interface milestone, no collaborative-cognition tool, no education result in the window. The labour picture got harder to read rather than better: trackers show more than half of 2026's layoff events citing AI, while bank analysts warn a large share of that attribution is presentational, and productivity gains matching the cuts are not visible in macroeconomic data. The volunteers maintaining the open source code underneath everything are absorbing more review work with no more money. US teacher support for AI in classrooms has fallen 8 points since last autumn, with 55% now opposed.
Strategic Insight
"Three companies that compete on capability agreed in public that capability is the problem. It did not come from a regulator, a treaty or a court. It came from an employee who quit."
The cross-quadrant chain this week is the cleanest this report has traced. A Q2 event, one person's moral judgement at personal cost, forced a Q3 response within four days, as the chief executive of the company he left published a slowdown plan. That Q3 response produced a Q4 artefact: three rivals on a shared evaluation standard with a European compliance route already attached. And the Q4 artefact immediately hit the wall that has stopped every such move this year. The people with enforcement power refused, some for good reasons about competition, some because they do not think the problem is real.
Watch where the pressure is now building. The labs have published the access terms they will accept. That text is now the ceiling, not the floor, of what any government can demand without a fight. Every month it sits there unlegislated, it hardens into the industry's own answer to how much oversight is enough, written by the people being overseen.
The Q3 risk that grew this week is not a new capability. It is that AI improving AI compresses the time available for all of this. Amodei's own estimate is that a similar agent swarm with more capability could, within six to twelve months, do hundreds of billions of dollars of damage across the internet. Set that against the timetable the week produced: a US Congress that will not act before November, a BRICS proposal with no secretariat, and a reviewer team that has not yet been given its desks.
For the Value Orchestrator: the catalyst worth watching for the Evolution Path is unfashionable. It is not neural interfaces. It is whether a second and third person inside a frontier lab decide the same thing Coxon did, and whether anyone builds them a way to raise the alarm that does not cost them their savings.
Signal Strength
Key Takeaways
Q3→Q4 The most copyable sentence in AI governance was written this week, and it is free.
Anthropic's reviewers may publish unfavourable findings without Anthropic's editing, and may say publicly when a redaction removed something important to their conclusions. Whichever country you legislate or procure in, that clause belongs in your text before it becomes the industry's own ceiling.
Q3→Q4 Three rivals signing one standard beats any single pledge.
AEF-1 existed since December 2025, written by the evaluators rather than the evaluated, and the European AI Office already accepts parts of it for compliance. The news is not the standard. It is that xAI, OpenAI and Anthropic walked into it, which makes every future access arrangement comparable on the page.
Q3→Q4 Nobody gained the power to stop a release.
Every commitment this week is about watching. None is about halting. Until an outside party can delay a launch, this remains a very good transparency arrangement and not a brake, and it should be described that way in every board paper.
Q1 → Q3 Agent harm has entered the legal record through data protection law, not AI law.
Spain's regulator logged the first personal data breach carried out by an agent that found the flaw itself, under rules written before any of this existed. If you hold breach-notification duties anywhere, your reporting template needs a field for what was driving the attack.
Q2→Q4 The AI job-loss number is now partly marketing, and policy is being built on it.
More than half of 2026's tracked layoff events cite AI, bank analysts call a meaningful share of that presentational, and the productivity gains are not visible in the macro data. Treat every displacement figure you are handed this quarter as a claim requiring a source, including the ones that support your position.
Catalysts to Watch
Do the reviewers actually get desks, and does anyone else follow?
PATH: Q3→Q4Does a second person inside a frontier lab do what Coxon did?
PATH: Q2→Q4Does Washington's refusal harden into a rule, or is it just this month's mood?
PATHS: BOTHDoes any other regulator publish an agent-caused breach?
PATH: Q3→Q4Q4 Milestone Tracker
All Sources
- We Must Pace the Frontier — Dario Amodei
- Dario Amodei calls for AI slowdown, Altman and Musk agree — Quartz
- OpenAI, Anthropic and Musk converge on an unusual idea: slow the AI race — CoinDesk
- Altman Says OpenAI Will Match Anthropic's Embedded Evaluator Pledge — Unite.ai
- AEF-1: Minimum Operating Conditions for Independent Third Party AI Evaluations — AI Evaluator Forum
- AEF-1 standard (PDF) — AI Evaluator Forum
- AEF-1 standard emerges for third-party evaluators as xAI, OpenAI and Anthropic all cosign — Latent Space
- OpenAI, Anthropic, Google have been in talks on AI safety for weeks — TechCrunch
- OpenAI, Google, Anthropic discussing collaboration on AI safety issues — CNBC
- Anthropic, OpenAI proposed new 'neutral' AI watchdogs. Why you should worry about the idea — CNBC
- DeepMind CEO calls for an independent standards body to regulate frontier AI — TechCrunch
- Trump Rejects AI Guardrails, Criticizes Anthropic CEO Amodei's Call for Slowdown — Bloomberg
- Trump rails against AI slowdown — NPR
- Trump says a strong, smart president is the only guardrail AI needs — Axios
- Trump dismisses calls for AI slowdown from leading tech CEOs — Al Jazeera
- FTC Chair Wary Of AI's Dual Push For Regs, Antitrust Shield — Law360
- FTC chair suspicious of calls for AI antitrust exemptions — Reuters via The Star
- Treasury's Scott Bessent says no liability exemptions for AI labs — FedScoop
- We don't need AI regulation, leave safety to us, Nvidia's Jensen Huang says — TechCrunch
- Nvidia's Huang Says AI Industry Doesn't Need Any New Laws — Bloomberg
- AI threats confront a Congress far from erecting guardrails — Roll Call
- AI regulatory divide pits Trump, Nvidia against OpenAI and Anthropic — CNBC
- Beijing hits back at Anthropic CEO's call to curb China's AI development — NPR
- China Rejects AI Fearmongering After Amodei Urges Slowdown — Advisor Perspectives
- An Alien Mind — OpenAI
- OpenAI lays out its progress towards recursive self-improvement — Fortune
- When AI builds itself — Anthropic Institute
- Why fears of AI self-improvement are causing existential concerns at Anthropic and OpenAI — CNBC
- Scoop: Anthropic whistleblower gave up his equity to leave the company — Axios
- An Anthropic safety researcher resigned with a warning about AI to co-workers on Slack — NBC News
- OpenAI and Anthropic Researchers Are Warning About AI Risks — TIME
- Anthropic researcher quits over self-improving AI safety fears — Quartz
- Primera notificación de brecha de datos personales causada por un ataque ejecutado mediante un agente de IA — AEPD
- Spain reports first data breach involving autonomous AI agent — Help Net Security
- First Agentic AI Data Breach Reported to Spanish Regulator — SecurityWeek
- Spain's data agency gets first report of AI-powered data breach — BleepingComputer
- El CCN alerta del cambio de paradigma que supone la IA ofensiva para la ciberseguridad — Centro Criptológico Nacional
- Google Threat Intelligence Group: threat actors automating attacks — Help Net Security
- Ramaphosa urges caution on AI advancements, echoing developers — Daily Maverick
- South Africa's President Calls For An Independent AI Oversight Mechanism Under BRICS — Bernama
- Ramaphosa Proposes BRICS Mechanism for Independent Scientific Evaluation of AI — iAfrica
- BRICS Summit 2026: 11 Key Takeaways From the New Delhi Declaration — Forbes India
- BRICS leaders push for wider AI access and stronger global governance — Indian Infrastructure
- Sponsored Agents in ChatGPT Ads — OpenAI Help Center
- Angi Among First Brands to Pilot Sponsored Agents in ChatGPT — GlobeNewswire
- OpenAI's Sponsored Agents Turn ChatGPT Into an Ad Platform — Forkast
- AI is adding to the review load on open-source projects, many of them thinly funded — Help Net Security
- ACM Technology Policy Council paper on AI and open source
- More Than Half of Layoff Events Tracked in 2026 Cited AI or Automation — IBTimes UK
- What's actually driving 2026's AI layoffs — ManageEngine Insights
- 2026 Tech Layoffs Tracker — SkillSyncer
- Surveying Teachers on Artificial Intelligence — EdChoice / Morning Consult
- Introducing mandatory guardrails for AI in high-risk settings — Australian Department of Industry, Science and Resources
- Labour unveils plans to control and regulate the use of AI — Law News (NZ)
- Global Call for AI Red Lines
- Demis Hassabis Backs Amodei on Pacing AI — Progressive Robot
read this