AI in 15 — August 05, 2026

A model that hasn't shipped yet, doesn't have a name you can buy, and might be called GPT-6 or might be called GPT-five-point-seven, just published ten proofs. And OpenAI declined to list a single human author on any of them.

Welcome to AI in 15 for Wednesday, August 5, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Today: the authorship fight over Astra, and the number that should actually unsettle you.

China's MiniMax opens the weights on a video model that beats everyone — then bans you from using it.

A worm ate eight hundred and sixty-eight npm packages overnight, and the security signal designed to stop it vouched for the malware instead.

Apple's lawsuit against OpenAI grows eleven names.

Plus Mistral ships a moderation model you can point at your own rulebook, Interpol puts a number on AI-enabled fraud, and a bubble argument worth taking apart.

Marcus, we've covered Astra for three days now. What's actually new this morning?

The authorship decision, which I don't think anyone properly digested over the weekend. OpenAI didn't just publish the proofs without peer review — they declined to name any human authors at all. Their stated reason is that claiming human authorship would, quote, misrepresent both the system's contribution and the nature of genuine human intellectual work.

That's a strange sentence to write about your own product.

It cuts both ways deliberately. And Thomas Bloom, who's been the most useful voice on this, pushed back on exactly that framing. He called the constructions big news — but his line is that the system was built by mathematicians and trained on everything mathematicians have ever written. So a paper with no human authors is a claim about the process that isn't really true.

What else did we learn about the model itself?

It's a family built for long-horizon work — coordinating multiple agents over hours or days, not a chatbot with a bigger context window. Altman demonstrated it to lawmakers on Capitol Hill on the twenty-ninth. Still no release date, still undecided whether it's GPT-6 or a point release.

And the two thousand dollars?

That's the figure that's travelled furthest, and it's worth being careful with. It's the cost of the tokens that produced the ten solutions, at current Sol API rates. It is not the cost of the search — we still don't know how many problems they pointed it at and got nothing. Noam Brown's own comment was, quote, sadly, no Millennium Prize problems yet, and that it's possible to push test-time compute much further.

So what's the honest takeaway?

If novel mathematics becomes a compute purchase rather than a career, the bottleneck moves from insight to problem selection. Choosing which question is worth two thousand dollars becomes the scarce skill. That's a real shift, and it doesn't require any of the ten results to survive peer review to be true.

MiniMax. Chinese lab, new video model, and Marcus, the headline says open.

H3. Thirty-three billion parameters, multimodal, and it's the first open model to top an Artificial Analysis video leaderboard. Number one in video editing, number two in text-to-video at twelve forty-two Elo — four points behind Google's Gemini Omni Flash, ahead of ByteDance. The technically novel bit is that it generates video and stereo audio in the same pass. Every prior system stitched two models together and dubbed the sound on afterwards.

So what's the catch?

Read the licence. Commercial use is permitted only for organisations under twenty million dollars in annual revenue. And per the reporting on the terms, it doesn't authorise use of the weights or their outputs in the US, the EU, the UK or South Korea.

Wait — so the four biggest Western markets are excluded.

Excluded. Which is an unusual thing to call open. And the download isn't the model that won the ranking either. Local deployment through ComfyUI tops out at 768p. The 2K module stays proprietary, and so does the component that turns a messy multimodal prompt into something the generator can actually use.

So you're getting the engine without the fuel injection.

That's about right. Hosted pricing is thirteen cents a second for 2K — roughly seven-eighty a minute, which MiniMax says is under a third of mainstream rates. ByteDance countered the same day with Seedance 2.5, closed, doing thirty-second clips with audio.

Is there real news underneath the licence, though?

There is, and it's this: the open-versus-closed gap has moved into video and audio, where the frontier had stayed almost entirely shut. That's genuinely new. I'd just stop calling it open weights and start calling it what it is — a distribution strategy with a geographic exclusion clause.

Security, and this one is ugly. Eight hundred and sixty-eight packages.

Yesterday, attackers took over the GitHub account of the maintainer behind keyv — a caching library with around a hundred and twenty-seven million weekly downloads. Same maintainer owns flat-cache, file-entry-cache, cacheable, cache-manager. Malicious commits went to the main branch, a release was cut immediately, and by lunchtime CEST there were eight hundred and sixty-eight packages across thirteen hundred and eighty-one versions confirmed compromised. Over two billion monthly installs in aggregate.

What's the payload?

A Mini Shai-Hulud variant. It steals credentials, then uses whatever publishing tokens it finds to poison further packages. Self-propagating. They also used an Ethereum smart contract as a resolver, so they can rotate infrastructure without touching the malware itself.

And there's a detail here that I think is the worst part.

It is. Because the releases went out through the normal GitHub Actions pipeline, the poisoned versions published to npm with valid provenance. The supply-chain integrity signal — the thing built specifically to tell you a package is legitimate — did exactly what it was designed to do and vouched for malware.

So the seal of authenticity authenticated the fake.

Provenance proves where a build came from. It has never proved the code is safe. Everybody knew that in theory. This is what it looks like in practice.

Where's the AI angle?

Two places. First — the same day, someone caught a Claude Code team-skills repository with malware committed to it, hidden files under dot-claude and dot-vscode designed to autorun. Agent skill files are instructions handed to a binary that has file-write and command-execution rights. That's the same trust decision as running a strange executable, with none of the accumulated caution.

Because they look like documentation.

They look like markdown. And second, the point one commenter made that I keep chewing on: a blackhat today would send agents to work every compromised repository automatically. The economics of supply-chain attacks change completely when the follow-on exploitation is itself automated.

Apple versus OpenAI got bigger yesterday.

Apple escalated the trade-secrets suit it filed in July. The new allegations: an ex-Apple employee met with an OpenAI engineer before a third person's OpenAI interview to discuss proprietary information on unannounced products. Another former employee allegedly took screenshots of confidential documents before interviewing. And Apple names eleven further ex-employees surfaced by its investigation.

Eleven. And OpenAI's response?

A point-by-point rebuttal titled "Apple is getting this wrong." They call the suit careless, aggressive and oddly personal, say it's based on false information, and that they don't have and don't want Apple's trade secrets. They posted emails they say show Apple staff asking their engineer for help finding files.

They're even arguing about the warning letter, aren't they.

Apple says it wrote in February and heard nothing. OpenAI says the letter went to the wrong person because Apple's lawyers confused two similar surnames. That's the level this is operating at.

What's the thing to actually watch?

The gap between the claim and the ask. "Some ex-employees mishandled documents" and "OpenAI must stop developing AI hardware" are very different propositions. Apple is seeking a preliminary injunction on the second using evidence for the first. And these are two companies that still need each other in the App Store, now litigating with receipts in public.

Mistral shipped something small and I think underrated.

Shieldstral. Three billion parameters, multimodal safety classifier, Apache 2.0 — genuinely permissive, worth contrasting with MiniMax half an hour ago. The design is the interesting part. Instead of emitting fixed categories, you hand it a plain-language policy at inference time along with the content, and it returns a single calibrated safety score. No retraining to change what you consider unsafe.

So the rules are an input, not a training run.

Exactly. Mistral says it matches or beats open guard models up to seven times its size, and it runs on a single sixteen-gigabyte consumer GPU. The broader trend it fits is labs shipping small focused models rather than stuffing everything into one frontier system. A dedicated moderation model is auditable in a way that hidden safety logic inside a general-purpose model simply is not.

Open question?

How wide the policy space really is. Can you point it at an arbitrary ruleset, or does it reliably reproduce only the one moderation style large platforms already use? Nobody's tested that yet.

Interpol has a number, and it's a big one.

Their African Cyberthreat Assessment for 2026, drawing on thirty-six member countries: fifty-five percent of reported cybercrime on the continent is now AI-enabled. Reported financial losses have more than doubled since 2024 — a hundred and ninety-two million dollars to four hundred and eighty-four million. Seventy-two percent of surveyed countries reported scam centres.

And the victims aren't local.

Largely not. Business email compromise has become markedly more convincing with generated text, and the operators are targeting people in Europe and North America. So this is a Western consumer-protection story as much as an African law-enforcement one. Operations are relocating from Southeast Asia and industrialising what used to be lone-operator fraud.

Do the defences scale the same way?

That's the whole question, and the answer is currently no. A fluent, personalised scam costs essentially nothing to produce now. Email and telephony still have effectively no sender authentication. One side got a hundred-fold productivity gain and the other side got a spam filter.

Last one. Ed Zitron's AI demand bubble piece, which got a lot of traffic this week.

His argument is about concentration, not capability. He reports OpenAI booked five-point-seven billion and Anthropic just under five billion in Q1 — together roughly eighty-nine percent of the largest AI companies' revenues. Against hyperscaler capex plans of over seven hundred and sixty-five billion this year.

And the dependency numbers?

Those are the sharp ones. OpenAI and Anthropic account for twenty-seven percent of Google Cloud revenue this year, projected forty-eight percent next. And sixty-nine percent of the year-over-year growth in Microsoft's Intelligent Cloud came directly from OpenAI. Strip that out, the segment grew eight percent.

How much weight do you put on it?

I'd interrogate it rather than accept it. The pushback is substantial — those two labs have annualised revenue around seventy-four billion and forty-nine billion from actual paying customers, growing fast, and if they vanished tomorrow the underlying demand wouldn't. Zitron has also predicted a crash several years running.

So what's the version of the question that's actually answerable?

Is the capex funding diverse demand, or building infrastructure whose only current tenants are two companies that lose money? That's a data question, unlike "is it a bubble." And there's a counterweight in today's numbers — a Clearwater survey of a hundred and seventy-eight fund managers found ninety-five percent raised AI budgets last year, eighty-five percent plan to raise them another fifty percent or more, and sixty-two percent are already running multi-agent orchestration in production.

That doesn't sound like two tenants.

It doesn't. Which is why the concentration figure and the adoption figure need holding at the same time.

One to watch: the White House frontier model framework. About a dozen companies met officials yesterday to close the loop on it, it's reportedly finalised, and the administration won't say what's in it. Astra would be the first model through.

Agreed — and the detail to watch is who gets to define "covered." Whoever writes that word writes the rules everyone ships under.

That's your AI in 15 for today. See you tomorrow.