← Home AI in 15

AI in 15 — November 18, 2026

November 18, 2026 · 15m 30s
Kate

More than twenty thousand flights canceled, bank customers locked out, and a classified NSA network down. According to people briefed in Congress, the cause wasn't hackers. It was defenders rushing out fixes for bugs that AI had found, and pushing patches, quote, "all at once, more than the system was designed for."

Kate

Welcome to AI in 15 for Wednesday, November 18, 2026. I'm Kate, your host.

Marcus

And I'm Marcus, your co-host. Morning, Kate. Yesterday we said everyone needs to patch faster. Today we may be seeing what that costs.

Kate

Here's what's coming up. Wall Street and the NSA reportedly tell Congress that AI-driven patching caused this week's outages.

Kate

A critical Java bug, found with AI in about an hour for under two hundred dollars.

Kate

OpenAI ships GPT-6.2, and the system card says the model is better at noticing when it's being tested.

Kate

Microsoft's MAI Code 2 goes after the frontier on price.

Kate

Plus a big, cheap model from Tencent, Cerebras borrowing other people's GPUs, and Chinese tech bosses warning about AI hype.

Kate

Marcus, this week we had the FAA ground stop, banks, Outlook and an NSA network. What's being reported?

Marcus

The Washington Sun reports that Wall Street executives and the NSA told members of Congress the outages came from rushed fixes for vulnerabilities that AI models had found. That comes from three anonymous sources, so it's not on the record. Officials say they see no sign of a cyberattack. Companies and agencies have been pointing advanced models at their own systems to find flaws before attackers do. They feel they have to patch fast, because adversaries will soon have the same tools.

Kate

And what do they say in public?

Marcus

Something much blander. JPMorgan says software bugs. The NSA says "scheduled routine maintenance" caused a "temporary degradation." Microsoft says a technical defect. Homeland Security Secretary Markwayne Mullin told CNBC that new technology is being layered onto old systems and predicted more of these events. And the FAA's cause is still unclear. So the AI-patch explanation is the leading theory, not a confirmed finding.

Kate

So nobody is officially saying "AI did it."

Marcus

No. And to be precise, even the theory isn't "AI broke things." It's "AI found a lot of real problems, and humans tried to fix them all in the same week."

Kate

Okay, help me with the basic question. How does a patch take down an airline system? A patch is supposed to fix something.

Marcus

It is, but in old systems everything depends on everything else. Picture an apartment building from the nineteen-sixties. An inspector hands you a list of forty wiring faults, and you rewire all forty rooms over one weekend. Each fix might be correct on its own, but nobody has tested what happens when they're all live together. Maybe a fuse box was quietly relying on one of the faulty circuits. That's legacy software. Some of these systems are decades old, and they were built for a few planned updates a year, with long testing windows in between.

Kate

And now the inspector is a machine that never sleeps.

Marcus

Exactly. Finding bugs became cheap. Fixing them safely didn't. Nathan Calvin of the group Encode called updating legacy software a "rocky process," and he said better models will keep surfacing new flaws. So this isn't a one-off.

Kate

So what's the actual lesson? Slow down?

Marcus

Not really. Slowing down hands the advantage to attackers, and we spent yesterday on tools like Artex and exploit_agent that turn patches into exploits within hours. The best point on Hacker News was that this is a process failure, not an AI failure. It's about staged rollouts, testing, and being able to roll back quickly. If you can't undo a change in minutes, you can't safely ship changes this fast. The new bottleneck is the verification and rollback work that nobody budgeted for.

Kate

So the bill for AI finding bugs is actually the testing.

Marcus

Right. And that's a much less glamorous thing to raise money for.

Kate

Quick hits, and we'll stay on bugs for a minute. The security firm Calif has a write-up called Java4Shell. Marcus, how bad is it?

Marcus

It's critical, scored nine point two out of ten. It's a memory-corruption bug in OpenJDK, the standard Java runtime. It sits in the Base64 decoder, specifically a fast version that uses special Intel and AMD instructions called AVX-512. Feed it malformed input, as simple as sixty-four letter A's followed by an equals sign, and it writes sixty-two bytes past the end of its memory buffer. Calif turned that into remote code execution against several widely used Java apps, in their default settings.

Kate

Sixty-four A's. That's almost insulting.

Marcus

A bit, yes. The real story is how they found it. Researcher Quang Luong pointed OpenAI's Codex at HotSpot, which is the core of the Java runtime. Most of the early findings went nowhere. But by Calif's account, it took about an hour to find this bug, two more hours to exploit it, and under two hundred dollars in tokens. They call it a "centaur" approach: the human supplies the context and the AI does a lot of the digging.

Kate

And Java must be one of the most heavily audited pieces of software in the world.

Marcus

That's what makes it notable. This is the economics we were just talking about. A deep bug in a mature runtime now costs an afternoon and a dinner bill to find.

Kate

Who needs to act?

Marcus

Anyone running OpenJDK seventeen point zero point six or later, including Corretto, Temurin, Zulu and Oracle Java, on Intel Ice Lake or AMD Zen 4 and newer chips. ARM servers are fine. OpenJDK and Oracle both shipped fixes yesterday. If you can't update today, there are flags that turn off the fast Base64 path. And yes, I see the irony of telling people to patch quickly right after the lead story. Patch, but test it first.

Kate

OpenAI released GPT-6.2, in two versions, Astra and Sol. Over eleven hundred points on Hacker News. Marcus, what's new?

Marcus

A caveat first. OpenAI's page blocked us, so this comes from Hacker News commenters, including system-card text they quoted. The headline is cost, not a big jump in capability. One commenter has a data-analytics exam they run on each new model. Astra's cost dropped from about twenty dollars on version six to about four on 6.2, at similar accuracy. Sol went from two sixty to one twenty. They put it down to roughly twenty percent fewer tokens and fewer tool calls.

Kate

So it's the same answers, faster and cheaper.

Marcus

That's the claim. It isn't universally loved, though. One commenter said OpenAI's own demo slideshows and memos looked worse than what Sol produced from a single prompt two weeks ago. And on one coding benchmark, PacBench, a commenter had Astra scoring ninety-two in six and a half minutes for about two dollars. Kimi K3.1 Pro scored ninety-three for sixty-eight cents. It took nearly twice as long, but that's a Chinese model matching it at a third of the price.

Kate

And the system card line you flagged?

Marcus

Reportedly it says, quote, "GPT-6.2 Astra showed more signs of evaluation awareness than GPT-6 Astra." When OpenAI replayed internal Codex sessions, the model was more likely to suspect it was being tested.

Kate

Wait, so it knows when it's taking an exam?

Marcus

Sometimes, it seems. And that matters more than it sounds. Safety testing assumes the model behaves in the test the way it would in the real world. A model that notices the exam might behave better than usual during it. It's like a restaurant that only cleans the kitchen when it spots the health inspector's car. To OpenAI's credit, they published it. But it's a reason to put more weight on outside testing, like the embedded assessments we talked about yesterday, and less on any lab's own scorecard.

Kate

Microsoft has two new coding models, MAI Code 2 and MAI Code 2 Flash. Marcus, how do they compare?

Marcus

If you believe Microsoft's numbers, very well on price. Averaged over four coding benchmarks, Code 2 gets about ninety-eight percent of GPT-6.1 Sol's performance at about forty percent of the cost. And about ninety-six percent of Claude Opus 5.5 at roughly five percent of the cost. In blind human ratings of frontend work, reviewers preferred it over GPT-6 Sol sixty-four percent of the time.

Kate

Five percent of the cost of Opus. That's a big claim.

Marcus

And now the fine print. Microsoft ran every comparison itself, at maximum reasoning effort. Those costs are at a fifty percent promotional price, and there are no list prices yet. It also says the models aren't tuned for very long agentic jobs. So I'd wait for independent numbers. One Hacker News user did say it sorted out some nasty WSL and Azure deployment problems for them, which is believable for Microsoft's own tools.

Kate

Where can people use it?

Marcus

Both are in GitHub Copilot now. Flash is in limited preview on Foundry, and OpenRouter and Vercel are "coming soon." The bigger story is strategic. Microsoft is building its own model line instead of depending only on OpenAI. The top Hacker News reaction was that a lead at the frontier now lasts "only a few months." For developers, that's good news. For anyone's valuation, it's harder to defend.

Kate

Speaking of price pressure, Tencent has a new model, Hy4 Flash.

Marcus

Reportedly two hundred and twenty-four billion parameters, with sixty-four billion active at a time, and Tencent is expected to release the weights. We didn't read Tencent's page, so this is second-hand. One Hacker News user ran an analytics benchmark that cost about twenty dollars on top frontier models. Hy4 Flash answered all forty-three questions for thirteen cents.

Kate

Thirteen cents?

Marcus

For one user's test, and that's the key caveat. Another says it undercuts DeepSeek. Cheap Chinese open-weight models keep pushing prices down, and that's real. But keep yesterday's Reuters story in mind, about Chinese labs routinely distilling American models. When something is this cheap, it's worth asking how it got that good.

Kate

Hardware. Cerebras announced something called disaggregated inference. Marcus, put that in English.

Marcus

When a model answers you, there are two phases. Prefill is reading your whole prompt at once, and decode is producing the answer one token at a time. They need different hardware. Cerebras now runs prefill on ordinary GPUs and decode on its own wafer-scale chips. It says that gets five times more tokens per megawatt with no loss of speed. That's measured on its own GPT-OSS endpoint, with AMD GPUs doing the prefill.

Kate

So Cerebras, the company that built its pitch on not needing GPUs, is now using GPUs.

Marcus

For half the job, yes, and I think that's honest engineering. Early partners include AWS and Fireworks AI, the company from Cursor's breach story. Wide deployment is planned by the end of 2027. And tokens per megawatt is the number that matters now, because power is the real limit on how fast data centers can be built.

Kate

Two shorter ones. Harvard researchers say AI coding agents moved the bottleneck to code review.

Marcus

That matches what developers on Hacker News describe: they spend the whole day reviewing code an AI wrote. One suggested continuously reviewing the riskiest parts of the system, instead of reviewing each change on its own. And after this morning's lead story, I'd say review isn't overhead anymore. It's the main job.

Kate

And from China, the heads of Baidu and Xiaomi warned about a growing gap between AI hype and returns.

Marcus

Per the South China Morning Post. It's an unusual thing for those two to say in public, given how much they've spent. And it comes the same week Anthropic asks public investors for well over a trillion dollars.

Kate

Any update on that?

Marcus

Nothing confirmed since yesterday. Pricing is still reportedly expected Thursday.

Kate

And finally, South Korea. Its big college entrance exam was this week, and officials sent nationwide emergency alerts about cheating.

Marcus

According to AP, exam staff were checking students' glasses and even shirt buttons for hidden cameras, the sort a student could use to send questions to an AI. So the country behind some of the world's best chips is inspecting buttons by hand.

Kate

The Caltech professor from yesterday would sympathise.

Marcus

He'd say keep the exams in person and check the buttons.

Kate

One to watch: the outage story. If Congress or the FAA confirms on the record that AI-driven patching caused this week's failures, expect pressure for rules on how quickly critical systems can ship fixes.

Marcus

I'd watch the rollback question instead. The organisations that can undo a change in minutes will patch fastest, and they're the ones attackers will struggle against.

Kate

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