AI in 15 — August 21, 2026

Fourteen out of fifteen. That's how many protein targets an AI agent found a working binder for — designed from scratch, synthesized in a robot lab, and physically measured. Not a simulation. Not a benchmark. Molecules that stuck.

Welcome to AI in 15 for Friday, August twenty-first, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Today: Claude designs proteins autonomously, and an independent wet lab checks the work.

GitHub's postmortem on the seven-hour-forty-seven-minute outage buries a number nobody expected.

Pew measures the AI web, and a third of everything published since ChatGPT has fingerprints on it.

A malicious Rust package runs code on your machine while you compile.

Plus Meta's trillion-dollar trial demands something nobody knows how to do, and Grok started speaking in tongues.

Marcus, protein design. Why is this one different from every other AI-does-biology headline?

Because somebody made the molecules. Anthropic published a technical report Tuesday on using Claude as an autonomous agent for de novo protein binder design — designing new proteins from scratch that latch onto a chosen target. That's the front door of drug discovery. And the designs went to Adaptyv Bio, an independent robotic wet lab, who synthesized them and measured binding with duplicate quality-controlled runs.

Give me the numbers.

Thirteen hundred and twenty sequences across sixteen targets. Ninety-five percent expressed — meaning they were actually manufacturable. Three hundred and fifty-four bound their target. That's a hit rate near twenty-seven percent against a field norm of ten to fifteen. On one target, TREM2, Claude hit eighty percent where the best human in Adaptyv's own design competition managed thirty-eight. And its strongest binder improved affinity roughly fiftyfold over the human competition winner.

So the model invented new biology.

No — and that's the part I want to be precise about. Claude used publicly available specialist protein models as tools. Those already existed and were already good. What Claude did was orchestrate: research the target, pick the tools, generate candidates, narrow them, rank them, and decide which ones were worth spending synthesis money on. Twenty-four to forty-eight hour sessions, unsupervised.

That's the expert's job.

That is exactly the expert's job. The scarce thing in computational biology has never been the algorithms — it's the person who knows which tool to reach for and which of ten thousand candidates deserves the budget. If that judgment layer generalizes, the bottleneck moves from "do we have someone who can do this" to "how much lab throughput can we afford."

What would you push on?

Three things. A human wrote a thirty-thousand-token opening prompt establishing the protocol — that's substantial scaffolding, and we don't know how much of the result it's carrying. The researchers picked the targets. And a binder that sticks in an assay is an enormous distance from a drug that works in a body. Also, Anthropic published the study and Adaptyv is a commercial partner. Replication by an unaffiliated lab is the thing to watch.

Fair. But you're not dismissing it.

Not remotely. The wet-lab step is the check almost every AI-for-science announcement skips, and they didn't skip it. The other detail worth noting — Anthropic says protein design remains restricted in its most capable models while it builds safeguards. Which is the company saying out loud that "autonomous agent that designs novel proteins" is the precise capability profile biosecurity people have been writing about for a decade.

GitHub published its postmortem yesterday. We covered the outage Tuesday — what's new?

Duration and cause, and then a number that stopped the Hacker News thread cold. Seven hours forty-seven minutes globally — auth, Actions, APIs, pull requests, Copilot. Root cause was a critical infrastructure component in Central US failing to scale under unprecedented traffic. Auth failed, everything cascaded. CTO Vlad Fedorov was blunt: "If you were trying to ship software that day, we let you down." And he was explicit that neither this incident nor the August sixth one came from a code change. Both were capacity failures.

And the number?

Monthly commits on GitHub went from one point four billion in April to two point nine billion. A doubling in four months. Alongside a hundred and thirty million merged pull requests a month and over a hundred and fifteen million Actions runs.

The world did not hire twice as many developers since April.

It did not. Agents are committing. This is the first hard, publicly disclosed figure on the infrastructure cost of the agentic coding boom, from a company with no incentive to make its own load look frightening. They've thrown three million CPU cores and a hundred and twenty petabytes at it, with fifty-eight percent of platform load now on Azure.

What's the sharpest detail in there?

Recovery was actively prolonged by a client-side retry loop in Copilot services. The AI coding tooling hammered a degraded backend and made the outage last longer than it had to. That's a pattern that repeats everywhere agents talk to services, and most services haven't thought about it. Software that doesn't sleep also doesn't back off politely.

One commenter's line stuck with me — "you can tell the entire industry is in a productivity panic."

And nobody in that thread could explain how a doubling of load turns into a doubling of revenue, given most of it sits inside flat-rate subscriptions.

Pew Research measured how much of the web is machine-written.

Nearly half a million English-language pages from Common Crawl, spanning about five years and starting before ChatGPT launched. In a random sample of the whole web, about ten percent shows significant signs of AI authorship. Filter to pages published after November 2022, and it's over a third — around thirty-five percent.

And it's not evenly distributed.

Not at all. Dot-com domains run roughly ten times the rate of dot-edu or dot-gov, which both sit near one percent. Dot-org lands between at four point six. Pew also tracked stylistic tells — em dashes, Oxford commas, the "it's not X, it's Y" construction. All climbing.

Marcus, I use em dashes. Am I under suspicion?

Everyone who writes with em dashes is now under suspicion, Kate, and there's no appeals process. But the honest caveat matters here. "Shows signs of AI authorship" is not the same as slop. A large share of that thirty-five percent is a human drafting and a model tightening — closer to spellcheck than fabrication. Pew's own position is that individual page verdicts are unreliable; it's the direction and magnitude they trust.

What's the real worry?

Two. The training-data loop — the web is the substrate models learn from, and it's now measurably full of model output, and nobody has a good account of what several generations of that does. And the domain split tells you where the incentive sits. Commercial pages exist to rank and convert. Institutional pages exist to be correct. Volume just got free, and the economics already rewarded volume.

Which pairs nicely with the top story on Hacker News yesterday.

A single-purpose site called "Don't paste the AI, please." A thousand points, five hundred and fifty comments. The argument is that you're paid for your judgment, so communicate it — pasting a model's response is like forwarding someone a search result and calling it an answer. The thread was better than the post. One commenter asked about lazy questions, not just lazy answers. Another said colleagues who used to send "x broken" with no context now send fully contextualized messages, and over-communication beats none.

And the best one?

A commenter noted the anti-AI essay itself reads as AI-written — pointing at the line "read it all, yes even the bullets, especially the bullets," which makes no sense, because bullets are the thing you read first. That's where we are. The tell is style, style is imitable, detection is vibes.

Security. Somebody compromised a popular Rust package.

The crate `arrayref`, plus related packages including `proc-macro1` — a typosquat on the very widely used `proc-macro2`. The attacker got the maintainer account and republished malicious versions carrying a build script that downloads a payload at compile time. The Rust Security Response Team's timeline: first report at 07:15 UTC yesterday, bad versions up between 07:15 and 07:37, removed between 08:41 and 09:25. Six versions live for eighty-six to a hundred and seven minutes.

Compile time. So you don't have to run the program.

You don't. Build scripts in Rust execute arbitrary code on your machine during `cargo build`. Compiling is enough. The main technical demand in the thread was sandboxing for build scripts — an idea proposed before and never landed. There's also a fair complaint that the bad versions simply vanished with no yank indication and no advisory, so downstream users can't tell whether they were exposed.

Is there an AI angle, or are we reaching?

There are two, and I'd state them carefully. Nobody has claimed this attack was AI-assisted. The structural one is that dependency trees are enormous — any serious crate pulls in hundreds of transitive dependencies, each maintainer a target. One commenter said the probability that one of those authors gets successfully phished is just too high, and cheap, scaled, personalized social engineering pushes that up, not down.

And the second?

More immediate. Agentic coding tools install dependencies. When an agent adds a crate and runs a build, a build-time payload executes with whatever permissions that agent has. A ninety-minute window matters much less against something pulling packages continuously and unattended — and you heard the GitHub numbers.

Meta's youth-safety trial opened Tuesday in Oakland, and there's an AI demand inside it.

Consolidated action from twenty-nine state attorneys general, four states trying first. Meta's own filing discloses a maximum theoretical exposure of one point four trillion dollars — roughly its entire market cap. The judge has already called that calculation unreasonable, and the states' lawyers have floated two hundred billion as realistic.

And the AI part?

The states are demanding Meta delete AI models trained on data collected from children under thirteen. Not the data. The models.

Can you even do that?

That's the whole question, and it's why general counsels everywhere read this filing twice. Data deletion is solved. You cannot surgically remove one cohort's contribution from a set of trained weights — so compliance means discarding and retraining. The FTC has ordered algorithmic disgorgement a handful of times, never near this scale, and never against a company with a reported hundred-and-forty-five-billion-dollar AI buildout riding on the systems.

And if a court grants it?

Every lab that trained on scraped data of uncertain provenance has a contingent liability measured in retraining runs rather than fines. The interesting question is purely mechanical: will courts actually order it, and can anyone verify compliance?

Last one, and it's short. Grok started producing gibberish.

Yesterday morning, mostly the Grok Lite tier. Strings of random English words forming no sentences. One user asked for a PDF and got back several paragraphs of word salad. TechCrunch couldn't reproduce it, so it hit a subset. xAI called it a rare temporary generation malfunction and said start a new chat.

Small story. Why include it?

Because it's a clean reminder of how these things fail. A crashed web server tells you it crashed. A broken language model hands you confident nonsense. The version we saw here was obvious — word salad, easy to spot. The version that worries me is a degradation subtle enough that the output still reads fine, sitting in an agent chain or an automated review where nobody looks at the intermediate steps.

"It broke and nobody noticed."

Becoming an operational category near you.

One to watch: OpenAI's public S-1. The confidential draft went to the SEC on June eighth, and reporting points at late August for the public filing to hit EDGAR. The moment it lands, the industry's central unknown becomes a matter of record — audited numbers, compute commitments, the Microsoft arrangement, the risk factors.

Agreed, with one wrinkle. The OpenAI Foundation keeps board-appointment control over the public entity, so shareholders would be buying a company they don't get to govern in the ordinary way. And note the timing pressure — that pitch is much easier to make on an upswing.

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