WEBVTT

1
00:00:00.000 --> 00:00:08.646
<v Alice>AI agents found an unknown flaw and went from mapping an app to writing files on its server in about 11 minutes. On their own.

2
00:00:08.946 --> 00:00:14.447
<v Bob>Plus a Riemann proof passing its first checks, and a very rough Friday for AI stocks.

3
00:00:18.747 --> 00:00:30.342
<v Alice>Welcome to Build Wiz AI Show. Big AI ideas, made clear for builders. I'm Alice, and this is our daily edition: the AI news of the last day, in about eight minutes.

4
00:00:30.642 --> 00:00:38.615
<v Bob>And I'm Bob. Everything today comes from The AI Gazette, this morning's edition and yesterday afternoon's. Six stories. Let's go!

5
00:00:38.915 --> 00:00:43.951
<v Alice>Story one is a security story. Bob, what did Amazon find?

6
00:00:44.251 --> 00:01:00.704
<v Bob>Amazon Threat Intelligence published findings on Friday about an attacker that ran Cairn, an open-source autonomous penetration-testing framework, against universities in East Asia. The attacker gave the agents goals, like obtain shell access, and let them plan and adjust on their own.

7
00:01:01.004 --> 00:01:03.912
<v Alice>Goals, not steps. That's new.

8
00:01:04.212 --> 00:01:23.637
<v Bob>Think of it as hiring a burglar and only saying: get inside. The burglar picks the window. Here, against an open-source medical imaging AI app, the agents found a path-traversal flaw and went from mapping the API to writing files on the server in about 11 minutes. Later, they got a shell running as root.

9
00:01:23.937 --> 00:01:25.147
<v Alice>Who's behind it?

10
00:01:25.447 --> 00:01:41.088
<v Bob>Amazon says a Chinese-linked group, with moderate confidence, based on Simplified Chinese prompts and China Standard Time timestamps. The Register identifies the software as MONAI Label, and Amazon says the maintainers have released a fix.

11
00:01:41.388 --> 00:01:43.951
<v Alice>And how did Amazon even spot them?

12
00:01:44.251 --> 00:01:48.939
<v Bob>The operator left the Cairn server publicly accessible, without authentication.

13
00:01:49.239 --> 00:01:53.717
<v Alice>Story two: Signal versus the AI assistant on your phone.

14
00:01:54.017 --> 00:02:12.312
<v Bob>Signal's president, Meredith Whittaker, wrote on Friday that the app will now restrict agentic systems built into iOS 27. On Android, it will block screen capture when it detects an AI assistant trying to read the Signal screen. She names Google's Gemini Intelligence and Apple's Onscreen Awareness for Siri AI.

15
00:02:12.612 --> 00:02:15.256
<v Alice>Why does Signal care what's on my screen?

16
00:02:15.556 --> 00:02:32.300
<v Bob>Imagine a friend reading your letters over your shoulder, then copying them into a notebook you never get to see. Signal says data captured from the screen ends up in whatever local database or remote server the agent uses, outside its control. It calls these integrations an existential threat to its mission.

17
00:02:32.600 --> 00:02:34.043
<v Alice>Can I turn it off?

18
00:02:34.343 --> 00:02:47.552
<v Bob>Yes, in settings. But Whittaker warns that could expose your messages, and your contacts' messages, to the agents. And Signal can't do the same on macOS yet, because the platforms give developers no easy way to restrict that access.

19
00:02:47.852 --> 00:02:51.947
<v Alice>Story three is about coding agents. Big money again?

20
00:02:52.247 --> 00:03:06.176
<v Bob>Very big. Bloomberg reports that Cognition, the maker of the Devin coding agent, is in funding talks at a 75 billion dollar valuation, two months after its last round. It's considering raising 2 billion dollars or more.

21
00:03:06.476 --> 00:03:10.281
<v Alice>Wait. In September it was 48 billion?

22
00:03:10.581 --> 00:03:20.006
<v Bob>Right. It's like a house whose price jumps every time the neighbours walk past. But these are talks only, from unnamed sources, and Cognition didn't respond to Bloomberg.

23
00:03:20.306 --> 00:03:22.358
<v Alice>And Cursor is in the news too?

24
00:03:22.658 --> 00:03:38.078
<v Bob>CNBC saw a report from two House committees saying Cursor plans to explore phasing out Chinese open-weight models because of geopolitical risk. Cursor built its Composer 2 coding model on Moonshot AI's Kimi K2.5. Explore isn't a decision, though.

25
00:03:38.378 --> 00:03:41.764
<v Alice>Story four is math. Big math.

26
00:03:42.064 --> 00:04:00.898
<v Bob>On Friday, Levent Alpöge, a mathematician who works at Anthropic, said he and Claude had found a proof of the Riemann hypothesis, formalized in the Lean proof assistant. The hypothesis, proposed in 1859, is about where the zeros of the zeta function lie, and it governs how prime numbers are distributed.

27
00:04:01.198 --> 00:04:02.955
<v Alice>So, is it true?

28
00:04:03.255 --> 00:04:22.669
<v Bob>Too early. But here's the update. Independent researchers, including MIT mathematician Andrew Sutherland, compiled the Lean proof themselves without issues, Scientific American reported. Think of Lean as a very strict referee who checks every single move, but only of the game you said you were playing. That's why a comparator run is still awaited.

29
00:04:22.969 --> 00:04:25.363
<v Alice>And there's a rival human proof?

30
00:04:25.663 --> 00:04:37.084
<v Bob>Stanford's Ravi Vakil posted a proposed all-human proof by postdoc Alessandro Masullo. About eight hours later, he added that several people had reported problems in it, found with the help of AI models.

31
00:04:37.384 --> 00:04:41.781
<v Alice>So now AI is checking the humans. What a reversal.

32
00:04:42.081 --> 00:04:46.304
<v Alice>Story five is the market. How bad was Friday?

33
00:04:46.604 --> 00:05:02.713
<v Bob>Bad. The Nasdaq dropped 5.3 percent on Friday, and Reuters put its weekly decline at 10 percent, the biggest since 2020. Nvidia fell 14 percent on Friday, after Reuters reported that the Securities and Exchange Commission is examining the company's accounting.

34
00:05:03.013 --> 00:05:04.868
<v Alice>Is that probe confirmed?

35
00:05:05.168 --> 00:05:12.885
<v Bob>No. Reuters reported it from unnamed sources, and neither the SEC nor Nvidia had confirmed it in the coverage the Gazette read.

36
00:05:13.185 --> 00:05:16.362
<v Alice>And Anthropic, just days after its debut?

37
00:05:16.662 --> 00:05:34.137
<v Bob>In its first full session, the stock fell about 9 percent, after being down as much as 14 percent at one point. Imagine the tide going out: it lowers every boat in the harbour, even the newest one. The Gazette's read is that the AI trade is being repriced, and Anthropic's debut didn't stop it.

38
00:05:34.437 --> 00:05:39.647
<v Alice>Last story: where AI gets built. What did President Trump say?

39
00:05:39.947 --> 00:05:51.833
<v Bob>At a White House press briefing on Friday, he said US tech companies building AI data centres in other countries would face a tremendous tax. He singled out Google's planned build in Finland, according to the BBC.

40
00:05:52.133 --> 00:05:54.070
<v Alice>How would that tax work?

41
00:05:54.370 --> 00:06:10.057
<v Bob>He didn't say. He gave no details on enforcement, and said he hadn't yet told his own officials. It's like a referee announcing a new rule in the middle of a match without saying what it is. So it's a spoken threat, not a policy: no order, rule or bill was published.

42
00:06:10.357 --> 00:06:13.350
<v Alice>Meanwhile, Oracle says it has its money?

43
00:06:13.650 --> 00:06:25.663
<v Bob>Oracle says it has secured all the funding it needs for its OpenAI contract, and calls itself the first hyperscaler fully funded for this investment cycle. Those are its own claims, and it gave no figure for the contract.

44
00:06:25.963 --> 00:06:29.165
<v Alice>Reality check. What should we hold loosely?

45
00:06:29.465 --> 00:06:39.481
<v Bob>Amazon's attribution is its own assessment, made with moderate confidence. The Riemann proof has no paper or peer review yet. And Cognition's round hasn't been announced.

46
00:06:39.781 --> 00:06:41.919
<v Alice>Fair. Hype check passed.

47
00:06:42.219 --> 00:06:45.212
<v Alice>Builder's Corner. One thing to do this week?

48
00:06:45.512 --> 00:06:59.917
<v Bob>If you run AI tools on a server, check that nothing is publicly accessible without authentication. An open server is exactly how Amazon caught these attackers. And if you use MONAI Label, install the fix. Lock it down. Then build.

49
00:07:00.217 --> 00:07:01.944
<v Alice>Today in three lines?

50
00:07:02.244 --> 00:07:16.173
<v Bob>One: autonomous agents are now an attacker's tool, not only a lab demonstration. Two: AI mathematics is moving from claims to checks. Three: the AI trade is being repriced, and where AI gets built is becoming political.

51
00:07:16.473 --> 00:07:20.406
<v Alice>What I'm watching: does the comparator run back the Riemann proof?

52
00:07:20.706 --> 00:07:25.195
<v Bob>And whether the SEC or Nvidia says anything about that reported probe.

53
00:07:25.495 --> 00:07:32.655
<v Alice>Today's stories come from The AI Gazette, this morning's edition and yesterday afternoon's. Thanks to them for the reporting.

54
00:07:32.955 --> 00:07:39.059
<v Bob>See you tomorrow morning. That's all for today on Build Wiz AI Show. Thanks for listening, and see you next time.
