AI in 15 — October 17, 2026

Thirteen years of conspiracy theories about a hidden NSA backdoor. The answer turned out to be a Unix timestamp from 1997 and a random number generator anyone can run. An AI model worked that out in under half an hour.

Welcome to AI in 15 for Saturday, October 17, 2026. I'm Kate, your host.

And I'm Marcus, your co-host. A cryptography mystery on a Saturday, Kate. Someone knows what I like.

I do. Here's what's coming up. Claude solves the twenty-five-year-old mystery of where the NIST elliptic-curve seeds came from.

Anthropic shows how it reviews thirty thousand code changes a day.

Postgres 19 is delayed because AI reviewers found more bugs than humans can fix.

Plus leaner quantum circuits, a map of what AI benchmarks actually measure, a new sandbox for agents, cheap open image models, and AI-written StarCraft bots.

Marcus, set this up for me. What are these curves, and why has anyone cared for twenty-five years?

In 2000 NIST standardized a set of elliptic curves for digital signatures. The best known is P-256. If you've loaded a website over HTTPS today, there's a good chance P-256 was part of that connection. Each curve is defined partly by a "seed," a number that's meant to be random. Those seeds came from the NSA, with no explanation.

And when the NSA hands you an unexplained random number...

Cryptographers get nervous. Especially after the Snowden documents and the Dual EC scandal, where an NSA-backed random number generator really did appear to be backdoored. The usual story was that an NSA mathematician, Jerry Solinas, made the seeds by hashing a joke passphrase. He said he couldn't remember the details, and he died in 2023. There's even been a public bounty for the answer.

So what did Anthropic's model find?

That they weren't hash outputs at all. Anthropic says an unnamed internal research model, given the problem without the hash hint, worked it out in under thirty minutes. It spotted that the top bits of some seeds were all zero. That's roughly a one-in-fifty-eight-hundred coincidence. From that it guessed the seeds came from lrand48, a plain random-number function in the C library. Then it brute-forced the hidden bits, used a lattice-reduction technique called LLL for the shorter seeds, and ran the generator backwards to its starting state. The starting state was a Unix timestamp from between May 1997 and April 1998.

So someone at the NSA basically typed "seed with the current time" and hit enter.

That's what it looks like. The model also traced six curves' base points to FreeLIP, a math library from the cryptographer Arjen Lenstra. And this is the part I like most: Anthropic published Python code that reproduces the seeds, released into the public domain. You don't have to trust them. You can run it yourself.

Hacker News loved it, I gather.

It was the number one story, almost eight hundred points. Thomas Ptacek called it "devastating news for an entire generation of NIST cryptography conspiracy theories." Another commenter wrote, "Thirteen years of tinfoil hats for srand48 of time of NULL."

Okay, now give me the Marcus caveat.

Two of them. First, this explains how the seeds were made, not how they were chosen. The generator produced candidates, and some valid ones were skipped. P-256 came from candidate number eighteen forty-two, when number eight seventy-six would also have worked. We can't see how many runs were thrown away, and Anthropic says plainly that it can't tell either. Second, a few things are still unexplained, including one binary curve seed that might just have a mistyped digit. And of course this is Anthropic reporting on its own model.

But nothing is broken.

Nothing is broken. If anything, a mundane, reproducible process makes a backdoor less plausible, not more. And the bounty, thirty-six thousand eight hundred and sixty-four dollars, goes to the Internet Security Research Group, the nonprofit behind Let's Encrypt. For me that's the real story: AI doing forensic research that anyone can check line by line.

Quick hits. On Wednesday we covered Anthropic replacing git. Now they've explained how they review all that code. Marcus, thirty thousand changes a day?

Up from about twenty-five hundred in May. And Claude now writes about ninety percent of Claude Code's own code. So they stopped pretending a human could read every diff. Their stated goal is "the minimum amount of human review time that doesn't raise reverts or incidents."

How does that work in practice?

Several AI bug-hunting agents review each pull request in parallel. Then verifier agents vote on each finding before it gets posted, which cuts the noise. A separate security pass can block the merge. And a classifier puts every change in a risk tier. Docs and low-risk changes need no human. Normal product code needs one. Billing, the auto-updater and the build system need two. And an agent's approval never counts as a human approval.

Does it work?

By their numbers, yes. Merged pull requests more than doubled from June to September, and the revert rate fell over forty percent. The share of PRs getting substantive review comments went from sixteen percent to fifty-four. And one pull request went through more than fifty overnight rounds of AI review and fixes before the author stopped it.

Fifty rounds. That's not a code review, that's a hostage situation.

Somewhat, yes. Humans keep architecture, scope and taste. And the rule is "you land it, you own it."

But I have to ask. Does Anthropic's software feel like a company merging thirty thousand changes a day?

That's almost word for word the sharpest Hacker News comment. Others said this is Google-scale process that only matters to Anthropic. My view: the volume numbers are the least interesting part. What matters is that review effort is matched to risk. Any team can copy that, whatever its size.

And now the flip side. PostgreSQL 19, the popular open-source database, is late. Marcus, blame the robots?

In a way. It was due around late September. Now it's expected about November nineteenth, and eight to ten significant features have moved to Postgres 20. The reason, according to The Register, is that AI-assisted code review keeps finding defects in code that was already committed, faster than volunteers can fix them. Core developer Peter Eisentraut put it well: "The effective quality bar has been raised after the time the feature was written and committed."

So the code didn't get worse. The inspection got better.

Exactly. Tom Lane said of the new graph-query feature that shipping it in 19 would mean bugs nobody could fix until 20. So they're waiting. Eisentraut still says 19 will be a great, very robust release.

And there's a bank version of this.

Business Insider reports Bank of America brought in about a thousand outside contractors, on top of roughly three thousand existing staff, to patch flaws that AI models found in its own and its vendors' software. Anthropic's Mythos is named among them. CEO Brian Moynihan put the cost, together with anti-money-laundering spending, at "a couple hundred million dollars plus" for the quarter. I couldn't read the full article myself, so take the details as reported. But the pattern is clear. Finding bugs got cheap. Fixing them didn't.

One Hacker News commenter said QA is becoming the real source of value.

I'd agree. Being able to fix things is now the bottleneck.

A second science story from Anthropic in one day. Claude has been designing quantum circuits.

In a closed loop with an automated checker, with no human input between rounds, Claude found designs for eight basic quantum computing building blocks that use fewer qubits and fewer gates than the best published versions. The standout is a circuit for elliptic-curve point addition with seventeen times fewer gates, at the cost of twelve percent more qubits. That took fifty-three hours of autonomous work, with proofs checked in Lean. Another, a decoder for a technique called decoded quantum interferometry, uses twelve times fewer gates than Google's version and cuts memory from as many as eight hundred and forty qubits down to nineteen.

Elliptic curves again. Is there a security angle?

There is. Elliptic-curve point addition is the core of a quantum attack on today's elliptic-curve cryptography. Cheaper circuits shorten the road to that attack. It doesn't make it possible tomorrow, but it's a reason to keep moving to post-quantum encryption. One caveat from commenters: some savings came from dropping parts that only handle very rare inputs, so in rare edge cases these circuits don't compute exactly the same thing. The work was led by Andi Gu with Harvard's Susanne Yelin and Mikhail Lukin.

Next, a reality check on "AI can do your job." Mercor mapped AI benchmarks onto the actual work Americans do.

They took twenty-nine benchmarks and matched them to the Labor Department's list of almost nineteen thousand tasks across nine hundred and twenty-three occupations. A task counted as covered if at least five benchmark prompts mapped to it. The result: ninety percent of American working hours go to tasks that benchmarks barely touch. Only two hundred and eight occupations have even one covered task.

And who is covered?

Mostly well-paid knowledge work. Lawyers are well covered. Legal secretaries and court clerks barely. So when a lab says its model can do a job, the evidence usually covers a narrow, high-paid slice of it. Mercor calls the numbers provisional. Commenters linked it to Karpathy's recent point that AI is really working on maybe two to five percent of all work.

So the honest version is "AI can do part of a lawyer's job, and we haven't checked the rest."

That's the fair reading.

Security. Yesterday we mentioned that Anthropic cut live internet access from its internal evaluations. Now Perplexity is giving away a sandbox for agents.

It's called SPACE. It's open source under Apache 2.0 and built on Firecracker microVMs. Those are tiny, fast virtual machines, the same technology behind AWS Lambda. Perplexity says nine frontier models tried to escape it more than a hundred times and none got out of the VM.

I hear a "but."

Four models did get past the network limits, through a flaw in the gateway's IP-based rules. It's been fixed. And honestly, that's the best possible advert for open-sourcing it. Outsiders can now poke at it too. CoreWeave, Akamai and Crusoe will offer it to customers. After the week of lab breakouts we've been covering, a free, tested, inspectable sandbox is exactly the boring tool the industry needs.

Quick follow-up on Meta's Muse Spark Open, which we covered yesterday. Apple has published real numbers.

Apple's machine learning team released a companion post with a new MLX version. On four M5 Ultra Mac Studios linked over Thunderbolt, the four-bit version, about four hundred and ninety gigabytes, generates about thirty-seven tokens a second at an eight-thousand-token context. Apple says peak power is about three hundred watts per machine, "a fraction of the power used by a common toaster."

Apple measuring AI in toasters. I love it.

The catch is longer context. At sixty-five thousand tokens, you wait about forty seconds for the first word. And commenters pointed out Meta's benchmarks tested Muse at four-bit against Qwen at full precision. My favorite line: "Meta will spend billions building open-weight models and then people will send Apple twenty thousand dollars to run them."

Staying with open models, Black Forest Labs has a cheap image model.

FLUX 3 Image Klein. Five point six billion parameters, open weights, native 4K output, editing with up to ten reference images, and it runs on an eight-gigabyte consumer GPU. About four seconds per image on an RTX 5090. The API starts under a cent per image. Krea 3, which we covered yesterday, is about six cents. It's free for companies under five million in revenue. Early users say it struggles with people, probably because of heavy filtering of its training data.

Something fun to close the quick hits. AI models writing StarCraft bots.

StarSkirmish Bench. Each model gets one hour to write a bot for the original StarCraft: Brood War, in C++, using its own lab's coding tool, with web access off. GPT-6 Astra came first. GPT-6.1 Sol came second and was the best value, at about a tenth of Astra's cost per run. Then Gemini 4 Argon, with Claude Opus 5.5 and Claude Fable 5.1 roughly tied behind.

And against human-written bots?

They get crushed. Astra wins only about seven percent of games against the top human-made bots. The interesting finding: the best predictor of a model's score was how often it practised against strong opponents. It barely tracks the usual intelligence rankings, which is exactly why it's worth having.

So the secret to AI StarCraft is the same as human StarCraft. Play better people.

That's how every teenager in 1998 learned it, Kate.

Lightning round.

An Anthropic Console user typed "5,04" in European number format and was charged five hundred and four dollars. An Anthropic employee confirmed the bug on Hacker News and promised a fix and a refund. Payment screens need a confirmation step. Viral GitHub pick: someone got llama.cpp running on a jailbroken PS5. It tops out around a twenty-seven billion parameter model at two-bit, and commenters are betting on a Sony takedown.

And money?

OpenAI's ads business reportedly has about sixty thousand advertisers, up from twenty thousand in late August, and a run rate around a billion dollars. 404 Media found advertisers can target users by uploaded email address and by what they type in prompts. That's worth knowing about. On guests: Geoffrey Hinton told The Guardian, "We're creating alien beings. I'm willing to choose humans." And Dwarkesh Patel argued on Tyler Cowen's podcast that a takeover by deceptive AI is plausible.

One to watch: Anthropic's IPO pricing. Reuters reportedly says investors are discussing about one and a half trillion dollars, down from two, and the FT reports a junk-grade credit rating from KBRA. The next few weeks will show what the market will pay for a frontier lab.

Agreed, but remember that KBRA isn't one of the agencies most bond buyers rely on. The prospectus will tell us more than the leaks.

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