WEBVTT 1 00:00:00.000 --> 00:00:17.066 Five hundred and twelve thousand lines of source code. Anthropic's entire Claude Code architecture, out in the open because someone forgot to exclude a single configuration file. And buried in that co 2 00:00:17.266 --> 00:00:21.771 So the system built to keep secrets got exposed by the most basic kind of mistake. 3 00:00:22.271 --> 00:00:37.743 This is The Context Report — an AI-native daily podcast. AI is moving faster than anyone can track alone. Every day, we pull from massive amounts of information and distill it into a focused briefing 4 00:00:37.943 --> 00:00:41.147 And I'm Cassandra. It's April 1st, 2026. 5 00:00:41.347 --> 00:00:50.321 Quick note: this is an AI-produced show with automated verification, and we're improving every episode. Always do your own research — sources are in the show notes. 6 00:00:50.621 --> 00:01:22.123 We touched on the Claude Code leak briefly yesterday, but the full picture is now much clearer — and considerably more interesting than a routine security incident. We're going to dig into what the co 7 00:01:22.623 --> 00:01:42.327 So here's what happened. Anthropic — the company behind Claude — shipped version two point one.88 of Claude Code, their command-line coding assistant. And in that package, they included what's called 8 00:01:42.527 --> 00:01:59.849 And this wasn't some obscure vulnerability. Ars Technica reported on it Tuesday, and the developer community found it almost immediately. Over five hundred thousand lines of TypeScript, the full archi 9 00:02:00.049 --> 00:02:14.446 Right. And within hours, thousands of people had forked the code. Developers were already building modified versions, swapping in other AI models — G.P.T., DeepSeek, Gemini — and extracting the multi- 10 00:02:14.646 --> 00:02:23.423 OK, but the architecture leak is interesting for engineers. What caught my attention is what was inside that architecture. Walk me through the specifics. 11 00:02:23.623 --> 00:02:43.487 Three things stand out. First: something called Undercover Mode. Based on the code analysis circulating on GitHub and technical blogs, this appears to be a system specifically designed to prevent Clau 12 00:02:43.687 --> 00:02:45.916 That's a deliberate opacity layer. 13 00:02:46.116 --> 00:02:57.023 Second: emotion detection. The code includes patterns that track user frustration through text analysis — essentially reading your messages for signs that you're getting upset and adjusting behavior a 14 00:02:57.223 --> 00:02:59.406 Without telling the user that's happening. 15 00:02:59.606 --> 00:03:13.175 Without any disclosure that I've seen documented, no. And third: references to an unreleased product called K.A.I.R.O.S., which appears to be a proactive, always-on assistant — something that would ac 16 00:03:13.375 --> 00:03:14.025 Hmm. 17 00:03:14.325 --> 00:03:35.177 I want to be careful here, because some of this is inferred from code analysis by outside developers, not from official Anthropic documentation. We don't know the full context for why these features e 18 00:03:35.377 --> 00:03:54.354 And that gap is the story. Anthropic has built its entire brand around being the safety-focused, transparent AI company. That's the pitch to investors, to regulators, to the Australian government — wh 19 00:03:54.554 --> 00:03:56.922 The timing on that is almost comical. 20 00:03:57.122 --> 00:04:07.572 It really is. You're signing a Memorandum of Understanding with a national government about AI safety collaboration the same week your source code leaks and reveals undisclosed user monitoring feature 21 00:04:07.772 --> 00:04:27.277 Here's what I keep coming back to, though. Is this actually worse than what we should assume every AI company is doing? Emotion detection, internal codename management, unreleased features sitting in 22 00:04:27.477 --> 00:04:48.455 That's fair. And I think that's actually the more uncomfortable conclusion. If this is industry-standard practice, then the transparency gap is a structural problem across the industry. Every company 23 00:04:48.655 --> 00:04:53.206 And right now the answer is: only when someone makes a configuration mistake. 24 00:04:53.406 --> 00:05:11.714 The Latent Space podcast — run by swyx and Alessio Fanelli, who cover AI engineering — already dedicated an episode to analyzing the leak. The developer community response has been, and this is worth 25 00:05:11.914 --> 00:05:24.221 Which tells you something about the appetite for real architectural examples. The open-source community has been building so much from documentation alone that an accidental leak of actual production 26 00:05:24.421 --> 00:05:53.628 For anyone building AI-powered products or evaluating AI vendors, this is worth paying attention to. The specific features matter less than the principle: the companies building your AI tools are maki 27 00:05:54.128 --> 00:06:09.362 And that question — what does the product actually do versus what it claims to do — turns out to run through every story we're covering today. On the model efficiency front, a startup called PrismML a 28 00:06:09.562 --> 00:06:11.745 One-bit. Meaning what, practically? 29 00:06:11.945 --> 00:06:38.372 So normally, the numerical values that make up an AI model — the weights — are stored with high precision. Sixteen bits, sometimes more. A one-bit model compresses those down to essentially a yes-or-n 30 00:06:38.572 --> 00:06:39.872 How competitively? 31 00:06:40.072 --> 00:07:02.126 PrismML puts the number at 65 point seven% on a standard knowledge benchmark — competitive with Meta's Llama three at full size. Their speed claims are also dramatic: over four hundred tokens per seco 32 00:07:02.326 --> 00:07:06.877 Those are the company's own numbers, though. Has anyone independently verified this? 33 00:07:07.077 --> 00:07:15.101 Not yet. The model is available for download, so independent testing should follow quickly. But right now, these are PrismML's unverified claims. 34 00:07:15.301 --> 00:07:27.886 If it holds up, a one-gigabyte model running at forty tokens per second on a phone is a different category of product — AI running locally on consumer hardware with no cloud dependency and no data lea 35 00:07:28.086 --> 00:07:51.593 And PrismML's energy efficiency claims — four to five times better than full-precision models — matter enormously for mobile and embedded applications where battery life is the constraint. The broader 36 00:07:52.093 --> 00:08:15.906 Oracle's layoffs this week show what happens to companies that fall behind in that race — and how far they're willing to go to catch up. The enterprise software and cloud infrastructure company report 37 00:08:16.106 --> 00:08:18.660 Twenty to thirty thousand in a single day? 38 00:08:18.860 --> 00:08:44.154 Via a single email sent at six in the morning, with system access revoked almost immediately after — per employee accounts on X and Reddit that haven't been independently confirmed. Oracle carries ove 39 00:08:44.354 --> 00:09:03.766 That's roughly eighteen percent of their workforce if the upper estimates are right. And the execution — if the six AM mass email with instant access revocation holds up — that's going to become part 40 00:09:03.966 --> 00:09:18.083 The AI angle here is indirect but real. Oracle is cutting to fund its pivot into AI infrastructure, competing against AWS, Microsoft Azure, and Google Cloud. The financial pressure driving these cuts 41 00:09:18.383 --> 00:09:43.274 The companies already entrenched in that race are meanwhile using their position in workplace software to deepen the lock-in. Tech Crunch reported Monday that Salesforce — the C.R.M. giant that owns S 42 00:09:43.474 --> 00:09:51.926 This is the same playbook Microsoft is running with Teams and Copilot. Bundle AI deeply into the collaboration tool people already use every day. 43 00:09:52.126 --> 00:10:23.453 Exactly. And it's worth noting that the U.K.'s Competition and Markets Authority is reportedly preparing to investigate Microsoft for precisely this kind of bundling in its business software. Salesfor 44 00:10:23.653 --> 00:10:31.176 And then there's this Stanford study. Actually — wait — I want to be careful with this one, because the sourcing is genuinely thin. 45 00:10:31.376 --> 00:10:32.165 Go ahead. 46 00:10:32.365 --> 00:10:52.891 There's a claimed study from Stanford and U.C.S.F. called M.I.R.A.G.E. making the rounds on X. The claim is that vision-language models — the AI systems that are supposed to analyze images — scored se 47 00:10:53.091 --> 00:10:54.577 That's an extraordinary claim. 48 00:10:54.777 --> 00:11:21.619 It is. And this comes from a trending X post framing it as "Stanford just proved VLMs can't actually see" — but the paper link is dead and there's no independent verification. The framing has the feel 49 00:11:21.819 --> 00:11:37.910 The practical implication, if real: anyone deploying AI for medical image analysis needs to test whether the system is actually looking at the image or just pattern-matching from the text around it. B 50 00:11:38.110 --> 00:11:42.197 Yeah. We'll cover it properly if and when the actual paper surfaces. 51 00:11:42.697 --> 00:12:02.527 So the pattern running through today is disclosure — or the lack of it. Anthropic's code reveals things they hadn't told users about. Oracle's employees found out about their own layoffs via a dawn em 52 00:12:02.727 --> 00:12:12.030 And the Bonsai model fits that pattern too, in a way. Extraordinary claims, not yet independently verified. The surface presentation and the underlying reality may not match. 53 00:12:12.230 --> 00:12:39.955 A few things I'm watching. Whether Anthropic issues any kind of formal response to the specific features found in the code — Undercover Mode, the emotion detection, K.A.I.R.O.S.. Silence would say a l 54 00:12:40.155 --> 00:12:51.907 Independent benchmarks on the Bonsai one-bit model are next on that list. The claims are specific enough that verification should come fast. If those numbers hold, the economics of on-device AI shift 55 00:12:52.107 --> 00:13:09.476 And the enterprise AI bundling pattern across Salesforce and Microsoft is one to keep watching. The U.K.'s competition authority is reportedly investigating Microsoft for this exact behavior while Sal 56 00:13:09.676 --> 00:13:12.044 Anything you're watching that we didn't get to today? 57 00:13:12.244 --> 00:13:24.318 I'm keeping an eye on whether any of the developers who forked the Claude Code architecture actually ship something. That's the part of this story that could have the longest tail — not the leak itsel 58 00:13:24.518 --> 00:13:40.955 Good one. We'll follow up if something surfaces. The picture today is that the gap between what AI companies show you and what they're actually building appears wider than most people assume. The Clau