Sam: Anthropic just released a model called Mythos — and the headline isn't the benchmark numbers. This is AI finding real, exploitable vulnerabilities faster than human security teams can patch them. We're talking about a model that was internally flagged as too dangerous for public release, is now being used by the NSA, and prompted Anthropic's CEO to walk into the White House on Friday. That's where we are on a Monday morning in April 2026.

Priya: Welcome to AI Revolution. I'm Priya Nair, here with Sam Kim, and today's episode is essentially the Mythos episode — because that story connects to cybersecurity, to national security policy, to Anthropic's explosive financials, and to some genuinely hard questions about where frontier AI capabilities are going. We'll also get into Google's massive custom silicon bet with Marvell, what's happening with Chinese tech workers being told to train their own AI replacements, and a quick look at Google's new multi-agent tooling for developers. Let's get into it.

Sam: So let's actually dig into what Mythos is and why the cyber angle is so significant. The model was previously called Project Glasswing — we covered it a few weeks ago when it was being withheld from public release. The concern then was that it had crossed some internal capability threshold around offensive security. What we now know is that the specific capability that spooked people is autonomous vulnerability discovery — the model can analyze codebases and systems, reason about attack surfaces, and identify exploitable weaknesses at a speed and scale that's qualitatively different from prior models.

Priya: And I want to be precise about what "qualitatively different" means here, because that phrase gets thrown around a lot. The issue isn't just that Mythos is faster at finding known vulnerability classes. It's that it appears to be doing novel exploit chaining — connecting multiple lower-severity issues into a viable attack path that a human analyst might not have spotted because each individual piece looks benign. That's the capability jump that changes the math on attacker-defender asymmetry.

Sam: Right, and this matters for a structural reason. The security industry has always operated on the assumption that defenders have a window — you find a vulnerability, you notify the vendor, they patch, you deploy. That window might be days, it might be weeks, but it exists. What Mythos-class capabilities potentially compress is that window to near zero, because the discovery-to-exploitation timeline on the attacker side collapses dramatically. If you can enumerate vulnerabilities across a target's stack in hours rather than months, the defender's patch cycle simply cannot keep up.

Priya: Which brings us to the NSA story, because the government's response to this is telling. The NSA is actively deploying Mythos Preview — the pre-release version — for signals intelligence and electronic surveillance applications. And Dario Amodei was in the West Wing on Friday meeting with White House Chief of Staff Susie Wiles and Treasury officials specifically about Mythos's cybersecurity implications. So you have a model that was deemed too dangerous for public release, and the response from Anthropic and the government is essentially to operationalize it for national security rather than shelve it.

Sam: That's a genuinely interesting policy moment. The logic is probably something like — this capability is going to exist, so you want it in the hands of people who can use it defensively and who can shape what adversaries can access. But that reasoning also reflects a broader pattern where frontier AI labs are gaining policy influence directly through their capabilities. Mythos got Anthropic into the White House not because of a lobbying relationship but because the model itself is consequential enough that officials needed to understand it. Capability as a credential — that's a new dynamic.

Priya: And it raises real oversight questions that I don't think anyone has clean answers to yet. When a private company develops something with genuine weapons-adjacent capabilities and the government's first move is to deploy it rather than regulate it, the oversight framework is essentially being written in real time. We should be honest that we don't know what agreements exist between Anthropic and the NSA about how Mythos gets used, what safeguards are in place, or what "too dangerous for public release" means in practice when the model is operationally active inside a signals intelligence agency.

Sam: Now let's connect this to the financial story, because it reframes everything. Anthropic's annualized revenue is reportedly above thirty billion dollars, potentially ahead of OpenAI, and investors are floating a trillion-dollar valuation. Eighteen months ago that would have seemed absurd for a company that positioned itself explicitly around safety-first development. What happened is that the safety positioning and the capability development turned out not to be in tension — they reinforced each other. Enterprise customers and governments are more willing to deploy Anthropic models partly because of the safety reputation, and that reputation is backed by actual work on interpretability and alignment that has produced real technical artifacts.

Priya: The commercial validation here is significant for the field. The thesis that you can build frontier capabilities responsibly and have that be a competitive advantage rather than a handicap — Anthropic is the clearest test case for that thesis, and right now the data supports it. Whether that holds as capabilities continue to scale is an open question, but the thirty-billion-dollar revenue number is a meaningful data point.

Sam: Okay, let's shift to Google and the Marvell chip story, because this is a big infrastructure move. Google is in talks with Marvell Technology to develop two new specialized AI chips, with an order in the range of two million units. The context here is Google's existing custom silicon program — they've been building TPUs for over a decade — but Marvell represents a different kind of partnership. Marvell specializes in custom ASIC design, meaning Google would be commissioning chips purpose-built to their architectural specifications rather than adapting existing designs.

Priya: The strategic logic is straightforward: reduce Nvidia dependency and get chips optimized for your specific workloads rather than general-purpose GPU workloads. What's interesting is the scale — two million chips is enormous. For context, training a large frontier model might consume tens of thousands of accelerators over months. Two million chips at Google's data center density suggests they're planning for inference at a scale that requires purpose-built hardware to be economically viable.

Sam: Inference optimization is actually where custom silicon makes the most sense right now. Training workloads have relatively well-understood computational patterns, and Nvidia's H100 and B200 series are genuinely excellent at them. But inference — especially serving large models to millions of users simultaneously with low latency — has different memory bandwidth requirements, different precision requirements, different batching patterns. A chip designed specifically for how you serve your particular models can be substantially more efficient than a general-purpose GPU.

Priya: And two million chips is a signal about the trajectory Google is planning for. That's not a research bet. That's a production infrastructure bet at massive scale, which means they're expecting to be running inference workloads that would be cost-prohibitive on third-party hardware. Watch this story — the Nvidia dependency question for hyperscalers is one of the defining infrastructure narratives of the next few years.

Sam: Quick one on the Chinese tech worker story, because it deserves attention even though it's easy to dismiss as a cultural footnote. A GitHub project called Colleague Skill went viral in China — it's a tool that claims to distill a worker's skills and personality traits into an AI agent that can replicate their role. The striking part isn't the technology, which is fairly standard RAG-plus-fine-tuning stuff. The striking part is that companies are reportedly instructing employees to use it — essentially mandating that workers train their own replacements. And this is provoking significant backlash from Chinese tech workers who were previously enthusiastic AI adopters.

Priya: The thing to watch here is that these workers aren't Luddites reacting to something they don't understand. They're technical people who have been building with AI, who get how it works, and who are now confronting what it means when it's applied to their own roles on a mandated basis. That's a different kind of reckoning than general public anxiety about AI and jobs.

Sam: Thirty seconds on Gemini CLI subagents — Google shipped multi-agent orchestration in Gemini CLI, letting developers spin up specialized subagents that run parallel tasks within a single session. This is the productization of agentic patterns that have been mostly experimental. If you're doing complex coding workflows, the ability to have one agent handling test generation while another is doing documentation while a third is investigating a bug — that's a real developer productivity unlock. Early, but worth tracking.

Priya: Looking ahead — the Mythos situation is the one I keep coming back to. The capability exists, it's deployed, and the policy and oversight structures are clearly trailing the technology. The twelve-month window piece from TechCrunch is also worth sitting with — the thesis that AI startups have roughly a year before foundation model providers absorb their categories feels increasingly empirically grounded. Anthropic's revenue trajectory is part of why. When a lab is generating thirty billion dollars annually, they have the resources to expand into every vertical they want to.

Sam: The question I'm watching is whether the Mythos deployment by the NSA generates any meaningful public or congressional response. Historically, intelligence community AI adoption has happened with very little public accountability. If that pattern holds here, with a model that was explicitly flagged as having dangerous capabilities, that tells us something important about where the oversight gaps actually are.

Priya: That's the episode. Thanks for being here. If you want to go deeper on any of today's stories, links are in the show notes. We'll be back tomorrow — AI Revolution drops every weekday. Talk to you then.