Sam: Anthropic has engineers sitting inside the NSA. About half a dozen of them, reportedly on-site, adapting the company's Mythos model for offensive cyber operations against China and Iran. The same company that publishes responsible scaling policies and is actively lobbying for a global AI pause button has its people embedded in signals intelligence, helping break into foreign networks. That's where we are on a Friday morning.

Priya: Welcome to AI Revolution for Friday, June 5th, 2026. I'm Priya Nair.

Sam: And I'm Sam Kim.

Priya: We have a packed show today. We're going to spend real time on the Anthropic-NSA story and what it reveals about dual-use AI policy. Then we'll talk about Anthropic's internal data showing Claude writing over 80 percent of their production code, and their simultaneous push for a global pause mechanism. We've got Microsoft caught misrepresenting how it trained its MAI models, Cloudflare saying bot traffic has already overtaken human traffic on the web, and some infrastructure stories that show just how frantic the compute buildout has gotten. Let's get into it.

Sam: So the Mythos-NSA story. Let me lay out what we know. Anthropic has stationed roughly half a dozen engineers directly at the National Security Agency. Their job is adapting Mythos — which is Anthropic's frontier model — for offensive cyber operations. The reporting indicates the model is being used to assist with network intrusion operations targeting infrastructure in China and Iran.

Priya: And the thing that jumps out to me immediately is the policy architecture that makes this possible. Anthropic's published acceptable use policies restrict certain surveillance applications, but those restrictions explicitly apply only to US citizens. That's the carve-out. If you're targeting non-US persons, the restrictions don't bind.

Sam: Right. And what's technically interesting here is what "adapting the model for offensive cyber ops" actually means in practice. You're probably talking about several things. One is vulnerability discovery — using the model to find exploitable weaknesses in target systems. Another is crafting payloads — generating code that can exploit those vulnerabilities in specific network environments. And then there's the operational planning layer, where the model helps map network topologies and suggest lateral movement paths once you're inside a system.

Priya: Each of those tasks individually is something frontier models have been getting better at. We've seen the benchmark results on CTF challenges and vulnerability detection. But deploying that in a classified operational context with engineers on-site to fine-tune for specific targets — that's a very different thing from scoring well on a security benchmark.

Sam: It is. And I think the reason this story hits so hard is the contrast with Anthropic's public positioning. This is the company that literally this week published data about why we might need a global AI development pause. They are simultaneously the most vocal safety advocate among frontier labs and an active participant in offensive state cyber operations.

Priya: I don't think those two things are necessarily contradictory, but they are in deep tension. And the way the policy carve-out is structured — protections for US persons, no protections for anyone else — that's a choice. That's not an oversight.

Sam: Let's pivot to the other big Anthropic story, because they're deeply connected. Anthropic released internal metrics showing that over 80 percent of their production code is now generated by Claude. Engineers are shipping eight times as much code per day as they were in 2024. And simultaneously, the company is formally advocating for a verifiable global AI development pause mechanism.

Priya: Walk through what 8x code output actually means technically, because I think people hear that number and either dismiss it or panic.

Sam: So what's happening is that the role of the engineer is shifting from writing code to specifying intent and reviewing output. You describe what you want at a higher level of abstraction, the model generates the implementation, you review it, iterate, and ship. The 8x number isn't 8x the lines of code — it's 8x the functional units being shipped. Features, fixes, integrations. The throughput of the engineering organization has increased by nearly an order of magnitude.

Priya: And the recursive element is the key part. Claude is writing the code that makes Claude better. Each improvement to the model potentially accelerates the next round of improvements.

Sam: Exactly. And I think that's precisely why Anthropic is pushing for the pause mechanism. They're seeing this acceleration curve from the inside. They know where it's heading. The pause proposal is conditional — they want it to be verifiable, meaning you'd need some kind of inspection or monitoring regime, and they want it contingent on other frontier labs participating so it's not unilateral disarmament.

Priya: Which brings us right back to the NSA story. The same acceleration curve that makes you want a pause button also makes your model more valuable to intelligence agencies. Those incentives pull in opposite directions.

Sam: They absolutely do. Let's talk about revenue for a moment because it contextualizes all of this. Anthropic's annualized revenue crossed 47 billion dollars in May. That's up from about 9 billion at the end of 2025. More than 5x in roughly five months. They're heading toward an IPO, so these numbers are going to face public market scrutiny.

Priya: 47 billion annualized is remarkable. That's the revenue trajectory of a company whose product has become infrastructure for a wide range of customers — from enterprises to, apparently, the NSA. It validates that frontier model capability translates to real spending at massive scale.

Sam: Okay, shifting gears. Microsoft has a different kind of problem this week. The company marketed its MAI models as trained on quote "enterprise grade, clean and commercially licensed data." That was a deliberate differentiator. Enterprise procurement teams care about IP liability, and Microsoft was telling them: our training data is clean, unlike the other guys.

Priya: And it turns out that's not accurate.

Sam: Investigation revealed that MAI models were trained partly on Common Crawl and other unlicensed web data. The same data sources Microsoft was implicitly criticizing competitors for using. Their actual approach relies on fair use doctrine and opt-out crawling — if a website doesn't explicitly block Microsoft's crawlers, the content gets scraped. That's identical to what every other lab does.

Priya: The technical distinction matters here. There's a real difference between a model trained exclusively on licensed data and one trained on a mix that includes Common Crawl. Licensed data gives you a clean chain of provenance. If an enterprise deploys a model and something in the output is traced back to copyrighted training data, the question of whether that training was authorized is material to the legal exposure.

Sam: And Microsoft specifically marketed around this. Enterprise customers who chose MAI models over alternatives partly based on this data provenance claim now have to reassess their risk posture. If you went to your legal team and said "we chose Microsoft because the training data is licensed," that conversation has to happen again.

Priya: It's a trust issue more than a technical one. The models probably work fine. But the procurement decision was made on a representation that turns out to be inaccurate.

Sam: Let's talk about the web itself for a minute. Cloudflare CEO Matthew Prince says bot traffic now exceeds human traffic on the internet. He'd previously forecast that wouldn't happen until late 2027. AI agents are the primary driver.

Priya: This is a structural shift. When more than half the traffic hitting websites is non-human, the economics of running a website change. You're paying for bandwidth and compute to serve AI crawlers, not paying customers or readers.

Sam: Prince's conclusion is that the web's future is "pay to crawl." Monetized data access rather than free scraping. And that has direct implications for model training. If you have to pay for every data source you crawl, training data costs go up significantly. That potentially advantages incumbents who've already trained on the open web and makes it harder for new entrants.

Priya: It also connects back to the Microsoft story. If pay-to-crawl becomes the norm, the distinction between licensed and unlicensed data becomes even more commercially important. You can't just lean on fair use if there's an established market mechanism for purchasing access.

Sam: Two quick infrastructure stories. AirTrunk, an Australian data center operator, is committing 30 billion dollars to build 5 gigawatts of AI data center capacity in India. That's among the largest single-country data center investments ever announced. It signals that the hyperscale compute buildout is globalizing fast — India is emerging as a major AI compute geography alongside the US and parts of Europe and the Middle East.

Priya: And then there's Meta, which is building data centers in tents. Literally temporary tent structures to house compute hardware. It's borrowing from Tesla's manufacturing playbook where they used tent structures to accelerate production timelines. The logic is pure speed — traditional data center construction takes 18 to 24 months, and Meta needs capacity now.

Sam: When a company worth over a trillion dollars is putting servers in tents, that tells you something about how urgently the hyperscalers view the compute race. Every month of delay is a month your competitors are training and you're not.

Priya: One more quick one. OpenAI updated ChatGPT's memory system. Instead of storing isolated bullet points about users, it now builds coherent narrative profiles organized by categories — work, hobbies, travel preferences. Memory accuracy improved from about 52 percent to 75 percent.

Sam: The technical shift is from key-value storage to something more like structured dossiers. It's more useful for personalization, but it's also the kind of persistent, categorized personal data accumulation that's going to draw regulatory attention, particularly under GDPR where the right to erasure and purpose limitation are pretty clearly defined.

Priya: Looking ahead, Sam — what are you watching after this week?

Sam: The Anthropic threads are what I keep coming back to. You have one company that is accelerating its own development with recursive AI-assisted coding, deploying its model inside the NSA for offensive operations, growing revenue at 5x year-over-year, pushing for a global pause button, and heading toward an IPO. Those threads are going to collide. When Anthropic is a public company, the tension between safety advocacy and classified government contracts becomes a disclosure question, a governance question, an investor question. I'm watching how that plays out.

Priya: I'm watching the data economics story. Bot traffic exceeding human traffic, pay-to-crawl emerging as a model, Microsoft getting caught on data provenance claims — these are all pieces of the same puzzle. The era of training on the open web for free is ending. How that transition happens will determine who can build the next generation of models and on what terms.

Sam: And the compute buildout stories — 30 billion in India, tents at Meta — tell you the industry expects demand to keep accelerating for years. Nobody commits that kind of capital unless they believe the growth curve continues.

Priya: That's our show for Friday, June 5th. Show notes and links to all the stories we discussed are at cleartext.fm.

Sam: Have a good weekend, everyone. We'll see you Monday.