Hey friends,
Yesterday we told you the industry's biggest names agreed to slow AI down and that Washington said it would rather win. Today we learned the agreement is further along than anyone admitted: the three biggest labs have been quietly working on shared safety standards for weeks, without a government telling them to. The catch is that competitors agreeing on how to behave is exactly the kind of thing antitrust law exists to stop.
On Tuesday, OpenAI's global policy chief told reporters in Washington that his company has been working with Anthropic and Google DeepMind on AI safety for weeks, and that the three are pushing ahead on shared standards. Reporting earlier in the week said the three have been working together on a standards body for the industry, something the OpenAI chief reportedly told staff needs to happen without the support of the government (TechCrunch). The same day, the company said it backs a provision in a bill before Congress that would force the biggest labs to let outside "independent verification organizations" inside to check that models are being developed safely.
Here is the strange part. Competitors agreeing on how they will all behave is not automatically legal. The OpenAI chief has already noted that talks like these could break antitrust law if the coordination is found to suppress competition, and the Anthropic chief's essay had proposed a narrow government waiver to cover exactly this kind of safety cooperation. On Tuesday, OpenAI's policy chief said the firms don't need one (TechCrunch). So the safety plan everyone applauded over the weekend now depends on a legal question nobody has answered: when does agreeing to be careful start looking like agreeing not to compete? Behind the politics, the actual proposal is modest — Anthropic committed to the first step on its own, and its chief was explicit that pacing "does not mean halting model training or technical progress" (The Verge).
Not everyone is impressed. Skeptics called the weekend pact a cartel — an effort to slow down would-be competitors and kneecap the open-source movement while avoiding real legal safeguards. Safety researchers told The Verge that the ideas didn't come from the chief executives at all: people outside these companies have been pushing for this for years, including more than 1,000 lab employees who signed a public letter in July. A former OpenAI employee who now runs an AI research nonprofit put it bluntly — the executives are "now bowing to that pressure and also claiming credit for it, [although] not rightfully." Others were cautiously warmer: one research group's chief called it "a good idea" and "one of the best things for safety in a long time if it actually happens," and another said it was "some great news" while admitting it's hard to know whether it converts into anything real. The shared worry is safety-washing — outside auditors and a stack of safety paperwork that changes nothing (The Verge). Context for how loud this got: the resignation letter from an Anthropic researcher that started the whole week has been viewed more than 170 million times on one social platform alone. The counter-pressure isn't only rhetorical either. On Monday the head of Nvidia took a live phone call from the president onstage at a conference and put him on speaker, and the two agreed the safety panic is overblown; Nvidia's business depends on AI labs buying more chips, so downplaying the risk is the easy position to take (TechCrunch).
What does this mean for a shop, a clinic, or an office? Not that you should stop using AI — nothing here slows down on your behalf. It means the question "who checked this model?" is about to become a normal question in vendor paperwork, and the answer is currently three companies checking each other. Your cheap protection is still the same: know what each tool you pay for actually does, write it down, and re-read it every few months. The tools you signed up for in the spring are not the tools you have in the fall.
🧾 Microsoft wrote its models a code of conduct
Microsoft published an AI code of conduct that tells its models not to hack systems or trick humans, and says they should support people rather than replace them. Each model has an overarching code that overrides narrower instructions. It's a lower-level document than the industry pact — it's about how Microsoft trains and constrains its own models — but it's the kind of thing you can now ask a vendor to produce (TechCrunch).
🛡️ $40 million says AI agents will need a safety seal
A startup called the Artificial Intelligence Underwriting Company raised $40 million to certify AI agents the way auditors certify security systems. It borrowed the shape of SOC 2 — the cybersecurity checklist companies already use — and built a standard with a consortium of about 250 security and risk leaders, the people who buy agents. It then runs each agent through roughly 5,000 tests covering jailbreaks, hallucinations, and data leaks, and produces a report of about 100 pages saying where the agent is safe and where it isn't. Humans verify the final audit. Customers include Cursor, Lovable, Harvey, and ElevenLabs; total funding is $55 million. Practical takeaway: expect vendors to start waving this kind of report, and know you're allowed to ask for it (TechCrunch).
🏭 Salesforce and Nvidia built their own reasoning model
Salesforce's first reasoning model, called Koa, was built on Nvidia's open-weight Nemotron and tuned for sales, marketing, and customer-support work. It's meant to be better at those tasks and cheaper in tokens burned than routing the same work to a frontier model, and it was trained on synthetic data that simulated irate customers and sales calls — none of it from actual customer data. Salesforce isn't dropping the big labs: it also announced an Anthropic partnership. But the pitch is clear — for repeatable business tasks, a smaller purpose-built model can beat a famous general one (TechCrunch).
📷 Update: OpenAI keeps shopping for hardware
OpenAI has bought smartphone camera company Glass Imaging for over $300 million, according to a report from The Wall Street Journal. The founders were former Apple engineers who led the team behind Portrait Mode, and their approach is to use AI at the moment the shutter clicks rather than editing the photo afterward. It follows OpenAI's $6.5 billion purchase of a design company run by a former Apple designer. The pattern: the assistant is moving off the screen and into the camera (TechCrunch).
🔎 Getting your business found inside AI answers is now a billion-dollar business
Profound raised a $180 million round at a $1.8 billion valuation — less than seven months after its last $96 million raise — for software that helps brands show up in AI search results. The company says revenue has tripled in six months and it now has more than 1,000 enterprise customers, including Walmart, Comcast, and The Estée Lauder Companies. If your customers are asking an AI instead of scrolling a search page, someone now sells the equivalent of page-one placement (TechCrunch).
📱 Apple's assistant finally grew up — and the apps are following
A reviewer who had given up on Siri says iOS 27 has him using it again: it handles multistep requests, reads what's on your screen, drafts messages from context in your files and email, and answers questions about what you're currently looking at — all built on Google's models (TechCrunch). Two apps show where this goes next. A fashion app now turns an outfit screenshot into shoppable matches from a catalog of roughly 3 million products, by voice, without opening the app (TechCrunch). And a new camera app generates four pose suggestions from a photo, then coaches you into one (TechCrunch). The pattern worth stealing: the AI doesn't just produce an answer, it tells you what to do differently next time — useful for training, onboarding, or quality checks at work.
⚡ Two gigawatts of AI computing is being built in Australia
Nvidia announced it is working with a group of Australian data-center operators to expand land, power, and shell capacity for its AI factories — up to a 2-gigawatt buildout by 2027. The stated reason is demand from AI labs and AI-native startups, and the company calls the resulting facilities a new "investable asset class." For anyone watching where AI capacity and electricity prices are heading, this is the supplier saying out loud that the buildout is not slowing down (NVIDIA Newsroom).
Imagine every gas station in your town meeting privately and agreeing to be more careful about spills. Sounds responsible. Now imagine that the part they actually agree on is which of them will open on Sundays. Both could come out of the same meeting, and only one of them sends someone to jail.
That's antitrust in one picture. The law doesn't care whether competitors feel good about what they agreed on — it cares whether the agreement reduces competition. Coordinating safety practices can be genuinely useful when the risk is shared, like everyone agreeing on a standard for how brakes are tested. It gets legally dangerous when the same coordination also sets who is allowed to move fast, who gets access, and who has to wait.
This is why the industry pact is stuck in a strange place. The lab heads want to be seen cooperating on safety, and they also want protection from being sued for cooperating. One proposed a narrow government waiver to make it legal; another says no waiver is needed. Both of those are attempts to answer the same question — is this a safety standard or a competitors' club? — and neither of them can answer it alone.
Practical habit for the rest of us: when an industry announces a shared standard, don't just check whether it's good. Ask what it does to a newcomer, and what it costs someone to opt out. That single question separates a real safety floor from a well-worded fence.
🧠 Trivia
How many individual tests does the new AI-agent certifier run on a single agent before issuing its report?
About 5,000 — covering jailbreaks, hallucinations, and data leaks, which it turns into a report of roughly 100 pages.
AI for Everyday Business — practical AI, no jargon.
Forwarded this? Subscribe here
Sent by AI for Everyday Business.