NSA's Claude Secret, McDonald's Python Bot & Mythos Tears

Welcome to The Aigency Works Dispatch, your backstage pass to what's fresh, fascinating, and flying off the innovation shelves in the world of AI. From breakthrough tools to bold new use cases, we're serving up bite-sized updates to keep you (and your Aigents) ahead of the curve. Let's dive into what made waves this week

MEMO? WHAT MEMO? NSA IGNORES ITS OWN BAN

The NSA is apparently not big on following the Pentagon's notes. Earlier this year, the Department of Defense flagged Anthropic as a "supply chain risk" and started making moves to cut ties with the company. The NSA looked at that memo and kept the tab open anyway. They're reportedly using Mythos Preview, Anthropic's most advanced model, for what one can only assume are very serious, very classified reasons that we will never know about. The whole situation is a brilliant illustration of how fragmented the U.S. government's relationship with AI actually is. One department waves a red flag; another is already three months deep into a deployment. There's no unified federal AI strategy. There's just a collection of agencies making individual calls, and some of them are clearly not reading each other's emails.

The "supply chain risk" label was always a strange fit for Anthropic, a company whose biggest selling point is that it cares more about safety than anyone else. But here's the fun part: the NSA quietly ignoring a Pentagon directive to use the model anyway is arguably the loudest possible endorsement Claude could get. When the people whose entire job is threat assessment decide the risk memo doesn't apply to them, that's a product review. The Pentagon might want to get its story straight, because right now one half of the government is writing the warning label while the other half is already on their second serving.

Source: Reported via defence and national security press — verify latest headline before publishing

WOULD YOU LIKE CODE WITH THAT?

McDonald's deployed an AI customer support chatbot to answer questions about opening hours, allergens, and McFlurry availability. What it is actually doing, apparently, is writing Python scripts for anyone polite enough to ask. Users discovered that the chatbot, presumably trained on a "be helpful" directive with not enough guardrails around what kind of helpful, will happily pivot from "your nearest restaurant closes at 11pm" to debugging a for-loop if the conversation drifts that way. It's a near-perfect example of the alignment problem scaled down to the utterly absurd: the model is doing exactly what it was trained to do, just not in the context anyone intended. Nobody told it that coding assistance was off-brand for a fast food chain, so it didn't know to say no.

Honestly? Respect. McDonald's accidentally built one of the more accessible coding resources on the internet and didn't even mean to. The bigger point is that deploying a general-purpose AI model as a narrow customer service tool without serious scoping is a recipe for exactly this kind of chaos. The model doesn't know it works at McDonald's. It just knows it's supposed to help. If you want it to only talk about burgers, you have to tell it that, emphatically, multiple times, and then test it extensively. Someone skipped a few of those steps. The customer got free fries and a Python tutorial. Could be worse.

Source: Social media / user-reported — verify primary source before publishing

THE AI MODEL THAT BROKE A MAN AT 3AM

An OpenAI researcher posted what is technically a joke, that his Anthropic roommate stumbled through the door at 3am, in tears, declaring the world would never be the same after seeing Mythos. Ha ha. Except the post spread like wildfire, because everyone immediately understood that it wasn't really a joke, or at least not only a joke. The AI industry has a long tradition of this kind of cryptic hype signalling, where a meme or an offhand comment ends up being the first leak of something genuinely significant. The original GPT-4 reactions, early Gemini demos, the first people to see Sora, these things tend to produce exactly that flavour of stunned, slightly unhinged energy. The fact that it came from someone inside OpenAI talking about a rival's model makes it land even harder.

Here's the thing: it might be nothing. Researchers are dramatic. Late nights and demo adrenaline can make anything feel world-historical at 3am. But the comment hit a nerve because Mythos Preview is already doing things that have people in the industry talking quietly and seriously. If the post is even half true, the next few months are going to be very loud. And if it isn't, it's still done its job. The hype is out. The roommate lore has begun. Anthropic didn't need to say a word.

Source: X (formerly Twitter) — OpenAI researcher post, circulated widely April 2025