Podcast 5-5 by Dongee · November 18, 2026

· 8:56 · Gabriel, Fernanda, Luis · Listen (MP3)

Today: a researcher using AI uncovers in one hour a Java flaw hidden for twelve years, and CISA tells everyone to patch right away; OpenAI launches GPT-6.2 Astra and Sol, and says Astra speeds up its own research, although an independent firm ranks it second behind Claude Fable 5.2; Nvidia posts the biggest report in its history, its first without Jensen Huang, and the board says it will not give him the job back; Waymo pulls off every freeway after California banned it over “imminent danger”; and Google will keep Gemini chats for about two months even if you turn off history.

The 5 stories

  1. A flaw had been hiding in Java for years, and a researcher using AI found it in one hour; the patch came out Tuesday, and CISA, the U.S. cybersecurity agency, told everyone to patch right away. It's in a piece that converts Base64 text — a way of turning files into text so you can send them over the internet: they made it faster for certain processors, and the bug is in that speedup. With text crafted on purpose — a row of letter A's and an equals sign, according to Calif, the cybersecurity company that found it — an attacker with no password could take control of the server. Calif says it tested the attack on widely used programs like Jira and Jenkins, that it cost less than $200 in AI, and that it took one hour to find the flaw and two more hours to build the test attack. The BBC reported it used GPT-6 Cyber, from OpenAI, which isn't open to the public. The patch came out the same day and only affects certain processors and versions; Calif is the same company that uncovered the PHP flaw in October, and nobody knows if there were any victims. An AI expert told the BBC it had been hidden in plain sight for twelve years. The flaw is known as Java4Shell (CVE-2026-114938). Source: Calif (disclosure and technical report), OpenJDK (vulnerability advisory, Nov 17, 2026), CISA (alert: patch immediately), BBC (GPT-6 Cyber; “hidden in plain sight for twelve years”), Datadog Security Labs (analysis), Hacker News
  2. OpenAI released two models: GPT-6.2 Astra, the powerful one, and Sol, the everyday one. An independent firm that compares models puts Astra in second place, right behind Claude Fable 5.2, but OpenAI says it beats Claude on a test of professional work with several steps, and that it does so at 5% of the cost; the company is the one reporting that test. For people who connect it to their own programs through the API, Astra is 20% cheaper than the previous model, and it's out for paying ChatGPT users and in Codex. OpenAI also brags that its documents “read less like AI.” Astra is, it says, its first released model to reach the “high” level of self-improvement, a measure of whether a model can speed up AI research; it doesn't improve itself, they compare it to giving every researcher a really capable assistant, and they say they're keeping an eye on it because that could move faster than the safeguards. Source: OpenAI (announcement), ChatGPT Learn (availability), The Verge, The Decoder (Artificial Analysis ranking), Hacker News
  3. Nvidia posted its first report without Jensen Huang, and the biggest in its history: $115.8 billion in revenue for the quarter, more than double last year and above what Wall Street expected. Colette Kress, who was the chief financial officer, is in charge on an interim basis, and for next quarter the company is forecasting around $130 billion, without counting any sales to China for data centers. That same Tuesday, Huang said he'd come back if they asked, and Harvey Jones, the board chair, said on the call with analysts that the board has no intention of giving Huang his job back. Almost 200,000 signatures ask for the opposite, according to CNBC. Huang resigned last week, and the search for a new boss goes on. Source: Nvidia (earnings release), CNBC (live; board, Kress, petition), CNBC (analysis), Yahoo Finance
  4. California's Department of Motor Vehicles talks about “imminent danger.” California pulled Waymo, Google's driverless taxis, off its freeways as of November 11, and Waymo suspended freeways in all its markets while it evaluates; they're still driving on regular streets. There were two incidents near San Francisco: three Waymos drove into an active construction zone with workers inside, and another tried to pass police cruisers that were holding back traffic, in a video that went all over the internet. The same thing happened in May. Waymo says there were no injuries and no crashes, and it stopped everywhere on its own, without being ordered to. The Verge says that without freeways, rides take longer and cost more. Source: The Verge, LAist (DMV statement), NBC Los Angeles, AFP, AP
  5. Google is splitting two things: saving your Gemini history and letting your chats train its AI. Now you can keep your history without training it. If you turn off history, your chats are kept for about two months; before it was 72 hours, according to TechCrunch. Google says that's to answer you and for security, with help from human reviewers, and it didn't explain what prompted the change. Temporary chats don't train the AI, and this rolls out in the coming weeks. ChatGPT can keep temporary chats for up to 30 days, and Claude, 30; Google, two months. Source: Google (blog), Gemini privacy hub, TechCrunch, 9to5Google

Transcript

Gabriel: Today on the Dongee podcast: AI helps uncover a Java flaw in one hour, OpenAI says its model speeds up its own research, and Nvidia posts a record, but without Huang.

Gabriel: Hey, hey. Wednesday, November eighteenth. Here with Fernanda.

Fernanda: Hi, Gabriel. A big hello to everyone listening. We're halfway through the week.

Gabriel: A flaw had been hiding in Java for years, and a researcher using AI found it in one hour. The patch came out Tuesday. Now it's a race.

Fernanda: A race between who and who.

Gabriel: The people patching against the people attacking. That's why the US cybersecurity agency told everyone to patch right away.

Fernanda: Hidden for years. How many.

Gabriel: An expert said that. I'm saving it for the end. First, Java. It's like the plumbing in a building: nobody sees it, but banks, airlines and telecoms all run through it.

Fernanda: And the leak is in one of those pipes.

Gabriel: It's in a piece that converts Base64 text. That's a way of turning files into text so you can send them over the internet.

Fernanda: About as basic as it gets.

Gabriel: Basic, but they made it faster for certain processors, and the bug is in that speedup. With text crafted on purpose, an attacker with no password could take control of the server.

Fernanda: No password. That's the scary part.

Gabriel: Calif, a cybersecurity company, found it. It says it tested the attack on widely used programs, like Jira and Jenkins.

Fernanda: And what does that crafted text look like?

Gabriel: A row of letter A's and an equals sign. That's it, according to Calif.

Fernanda: Another row of letters, like the BugMeNot one. At least this one comes with an explanation.

Gabriel: And the bill: less than two hundred dollars in AI, says Calif.

Fernanda: Dirt cheap.

Gabriel: One hour to find it and… well, two more hours to build the test attack.

Fernanda: You still haven't said which AI.

Gabriel: The BBC reported it: GPT-6 Cyber, from OpenAI, which isn't open to the public. And if the good guys have something like that, the bad guys are going to go looking for their own.

Fernanda: The patch came out the same day, Gabriel. And it only affects certain processors and versions. Plus, Calif makes its living from this. It's the same company that uncovered the PHP flaw in October.

Gabriel: And nobody knows if there were any victims. But here's what I saved for you: twelve years. Hidden in plain sight, an AI expert told the BBC.

Fernanda: Twelve years, and one hour to find it.

Fernanda: OpenAI is putting out new models. An independent firm that compares models puts the powerful one in second place, right behind Claude Fable 5.2. But OpenAI says it beats Claude.

Gabriel: Beats it at what?

Fernanda: Hold on, first what came out. There are two: GPT-6.2 Astra, the powerful one, and Sol, the everyday one.

Gabriel: Two at once.

Fernanda: And there's a line in the announcement I had to read twice. That's coming at the end.

Gabriel: What I did see is the price. For people who connect it to their own programs, through the API, Astra is twenty percent cheaper than the previous model.

Fernanda: And it's out for paying ChatGPT users…

Gabriel: …and in Codex. I pay for it, and I use it for cooking recipes.

Fernanda: A cutting-edge model for your mac and cheese.

Gabriel: Okay, now: what does it beat it at?

Fernanda: A test of professional work with several steps. OpenAI says it beats Claude at five percent of the cost.

Gabriel: Five percent. That's practically free.

Fernanda: If it's true. The company is the one reporting that test.

Gabriel: Noted.

Fernanda: And it brags about something else, that its documents "read less like AI."

Luis: I write like a robot with no help from anyone. My emails read like a washing machine manual.

Gabriel: That's why nobody writes back, Luis.

Fernanda: The line I owed you. Astra is its first released model to reach the "high" level of self-improvement, says OpenAI. That measures whether a model can speed up AI research.

Gabriel: Wait. It improves itself?

Fernanda: No. They compare it to giving every researcher a really capable assistant. And they're keeping an eye on it, they say, because that could move faster than the safeguards.

Gabriel: More capability for a lot less money opens doors for small teams, in places like ours too. Then again, the one measuring, watching and announcing is the same company.

Fernanda: Same company, yeah. And it came in second.

Gabriel: Nvidia's first report without Jensen Huang, and the biggest in its history. That same Tuesday, Huang said he'd come back if they asked. The board said no.

Fernanda: A record and a slammed door, same day.

Gabriel: The board part comes later. First, what everyone was waiting for on Tuesday: revenue of one hundred fifteen… one hundred fifteen point eight billion dollars for the quarter.

Fernanda: More than last year, I'm guessing.

Gabriel: More than double. And above what Wall Street expected.

Fernanda: Without the founder. And with someone interim in charge.

Gabriel: Colette Kress, who was the chief financial officer. And for next quarter, it's forecasting around one hundred thirty billion.

Fernanda: Counting China?

Gabriel: Without China. It isn't counting on any sales to China for data centers.

Fernanda: Okay, now the board.

Gabriel: Harvey Jones, the board chair, said it on the call with analysts: the board has no intention of giving Huang his job back.

Fernanda: And outside, people are asking for the opposite. That whole petition.

Gabriel: Almost two hundred thousand signatures, according to CNBC.

Fernanda: That's a lot of people.

Gabriel: Huang resigned last week, and the search for a new boss goes on.

Fernanda: No boss, and a record.

Gabriel: The numbers say the company works without its founder. Demand for AI chips is still huge.

Fernanda: But they're looking backward. Who runs it, the open research, China left out of the math. A record doesn't fix any of that.

Fernanda: "Imminent danger," says California's Department of Motor Vehicles. Waymo, Google's driverless taxis, says there were no injuries and no crashes. And they still pulled off every freeway.

Gabriel: California's freeways, or every single one?

Fernanda: California pulled them off its freeways as of November eleventh. And Waymo suspended freeways in all its markets while it evaluates. They're still driving on regular streets.

Gabriel: They must've done something to earn that.

Fernanda: Two incidents near San Francisco.

Gabriel: Go ahead.

Fernanda: The first: one Waymo… no, three, drove into an active construction zone…

Gabriel: …with workers inside. And here I thought the Cybercab driving off the map was the weirdest thing this week.

Fernanda: The second one has video, and it went all over the internet. Another Waymo tried to pass police cruisers that were holding back traffic.

Gabriel: I saw that video.

Luis: Never driving again, that was my plan. But not in a car that passes the cops.

Gabriel: You can't even pass a bicycle, Luis.

Fernanda: And it's not the first time.

Gabriel: When else.

Fernanda: The same thing happened in May.

Gabriel: Still, nobody got hurt. And Waymo stopped everywhere on its own, without being ordered to. They don't take a human driver off the freeway for that.

Luis: I'd ride on the freeway with a robot, no problem. With my cousin driving, no way.

Fernanda: That's you, Luis. My dad spent years fixing roads. I think about the guys in the vests first.

Gabriel: Them first, sure. But anyone who uses Waymo is going to feel it: The Verge says that without freeways, rides take longer and cost more.

Gabriel: Do you have your Gemini history turned off, Fernanda?

Fernanda: Since day one.

Gabriel: Well, there's good news, and there's one that hits you. And at the end, a comparison with ChatGPT and Claude.

Fernanda: Good news first.

Gabriel: Google is splitting two things: saving your history and letting your chats train its AI. Now you can keep your history without training it.

Fernanda: A lot of people were asking for that. Now the one that hits me.

Gabriel: If you turn off history, your chats are kept for about two months. Before it was… let's see, seventy-two hours, according to TechCrunch.

Fernanda: Two months. With it turned off.

Gabriel: To answer you and for security, says Google, with help from human reviewers. What prompted the change, it didn't explain.

Fernanda: And with human reviewers.

Gabriel: I've asked Gemini things I wouldn't even ask my doctor.

Fernanda: So some reviewer knows more about your knee than your doctor does.

Gabriel: Here's what is clear: temporary chats don't train the AI, and this rolls out in the coming weeks.

Fernanda: More control, I'll give them that. But the two months are still there.

Gabriel: Now, the comparison: ChatGPT can keep temporary chats for up to thirty days, and Claude, thirty. Google, two months.

Gabriel: That's it for today's episode. Tomorrow, five more stories.

Fernanda: And Luis, signing his washing-machine-manual emails with a row of letter A's.

Gabriel: Talk to you tomorrow.