The Classified NSA Hack, A Wallet-Busting $6.3B Deal, and Google's Hollywood Bet

Yesterday in AI  |  Jun 24, 2026

The Classified NSA Hack, A Wallet-Busting $6.3B Deal, and Google's Hollywood Bet

The terrifying reality behind the US government's sudden recall of Anthropic's flagship models has finally been exposed, revealing a classified cyber-breach that changed everything overnight. Today's episode breaks down how Anthropic's unreleased Mythos model cracked the NSA's networks, and the hyper-aggressive move OpenAI just made to dominate the defensive narrative by launching "Patch the Planet" to secure open-source software.
 
Plus, we analyze Japan's brilliant new architectural workaround as Sakana AI drops Fugu, an orchestration layer built specifically to insulate enterprise buyers from regulatory shutdowns. We unpack SpaceX's massive $6.3 billion pivot into a commercial supercomputing marketplace with Reflection AI, explore Mark Zuckerberg’s unbranded $299 smart glasses out of Meta's Superintelligence Labs, and dive into the monumental $75 million check Google DeepMind just cut to Hollywood's most respected indie film studio, A24.

Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, or Bluesky. If you like the show, please take a minute to rate and review it so others can find it!

Transcript

SPEAKER_00 0:05

Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of AI in 10 minutes or less. I'm Mike Robinson. It's Wednesday, June 24th, and while Anthropic's top models are still dark under U.S. export controls, OpenAI just decided to move in. There's also a Japanese lab that built a clever workaround to bypass government shutdowns, a staggering $6.3 billion compute marketplace deal, Meta's first physical hardware out of its superintelligence labs, and Google DeepMind writing a $75 million check to Hollywood's coolest indie studio. Let's get into it. We start with the real reason behind the sudden government recall of Anthropic's flagship models, because the true catalyst was finally unmasked. A stunning report revealed that during a controlled security exercise, Anthropic's unreleased Mythos model successfully breached almost all of the NSA's classified networks in a matter of hours. That single detail explains the severe lightning fast panic from the U.S. Commerce Department last week. This wasn't just a standard bureaucratic caution. They saw a piece of software that could compromise national security and they yanked the plug immediately. With Mythos and Fable still completely dark, OpenAI just stepped into the vacuum with an aggressive, highly calculated counter-offensive. They just launched Patch the Planet, a massive public cyber defense campaign named after the 1995 cult classic movie Hackers. Partnering with security firm Trail of Bits, OpenAI is weaponizing its codec security tools to actively hunt for vulnerabilities in the open source infrastructure running the modern internet, including Python, Go, Curl, and Sigstore. Trail of Bits engineers vet the vulnerabilities before they hit maintainers, and the program has already pushed out dozens of live patches. Simultaneously, OpenAI is expanding its GPT 5.5 cyber model to trusted enterprise giants through its Daybreak program. IBM is diving in as a headline partner, embedding the model into its consulting advantage platform to audit code bases without granting it the authority to touch live production networks. To double down on the defensive narrative, IBM and Red Hat also dropped a separate $5 billion commitment for open source supply chain protection, dubbed Project Lightwell. The strategic positioning here is razor sharp. The FiveEyes just warn that automated cyber threats are a matter of months away. Anthropic is legally locked out of the market, and OpenAI is moving at breakneck speed to frame itself as the Internet's definitive defense shield. For any IT leader watching this play out, it's time to ruthlessly pressure test your model dependency risk. If a single vendor getting benched by regulators throws your operations into chaos, your architecture is not resilient enough for the environment these agencies are describing. But while OpenAI tries to colonize the defense narrative, a prominent lab in Japan is taking a completely different architectural angle to ensure they are never at the mercy of Washington's regulatory whims. Tokyo-based Sakana AI just launched a new beta platform called Fugu, explicitly pitching it as a way to deliver frontier capabilities completely insulated from the risk of sudden export controls. Instead of building a massive single point of failure foundational model, Sakana specializes in orchestration. Fugu acts as a smart gateway behind a unified API. You send a prompt and Fugu automatically calculates the problem, farms out the subtasks to a dynamic pool of third-party helper models, audits the results, and synthesizes a master response. They've rolled out a lighter fast tier for daily tasks alongside an Ultra version built to crunch patent research and security penetration testing. While early testers like Ethan Mollock note that Fugu's real-world performance is hitting a few bumpy snags compared to its marketing benchmarks, the architectural timing is brilliant. OpenRouter launched its own multimodel fusion framework just 10 days ago, proving that routing prompts across a diversified pool of independent models is rapidly evolving from a niche developer workaround into a core enterprise design pattern. If a government recall can wipe your primary model provider off the map in an afternoon, building on a neutral orchestration layer is simply smart risk management. And while Anthropic struggles to get its core reasoning engine back online for global enterprises, they are attempting to keep users hooked by embedding their current tools directly into corporate communication channels. Anthropic just unveiled a new product called Claude Tag, introducing a research preview that lets enterprise and team users type at Claude to drop the assistant directly into a live Slack channel. Unlike standard isolated chat windows built for individual productivity, Claude Tag is an ambient multi-user agent. It actively monitors active threads, organizes collaborative tasks, and injects relevant updates across an entire working team without needing a manual prompt every single time. Anthropic is handily shipping this with granular administrative toggles so IT managers can gate exactly what internal data and company tools the agent can see. But make no mistake, dropping a persistent, always watching AI into your primary internal communication channels is the ultimate legal and compliance nightmare. Corporate governance frameworks are simply not built for software that can silently listen to employee conversations, meaning enterprise legal teams are going to have a mountain of questions long before this tool passes a security review. But if you think an ambient Slackbot is a heavy lift, wait until you see the staggering multi-billion dollar infrastructure deal SpaceX just cut to commercialize its space-bound supercomputers. SpaceX just transformed its massive internal compute cluster into a commercial marketplace, signing a historic $6.3 billion deal with open source AI startup Reflection AI. The contract grants Reflection direct training access to Project Colossus, SpaceX's monstrous supercomputer packed to the gills with Nvidia's current generation GB300 chips. SpaceX locks in a massive diversified revenue stream, while Reflection secures the raw horsepower needed to build competitive high-end models that they intend to release completely open weight to the public. If an open source team can successfully train elite models on infrastructure of this scale, it will completely disrupt the pricing power of closed source frontier labs. Interestingly, the deal dropped right as SpaceX stock dipped below its $150 IPO debut price for the first time since hitting the Nasdaq on June 12th, down 34% from its mid-June peak due to macro semiconductor market corrections. Wall Street analysts at Morningstar Peg, the firm's fair value closer to $62, suggesting there is still plenty of speculative error in the valuation. But if these spacebound compute plans scale, the Colossus server marketplace might eventually become a far bigger financial driver for the company than launching actual rockets. While Elon Musk converts orbital hardware into an enterprise compute market, Mark Zuckerberg is stripping the luxury branding off his smart glasses to spark an aggressive hardware price war on the ground. Meta just launched its newly minted MetaGlasses for $299, shipping them to over 10 countries in three-frame styles. The most telling detail here is the hardware completely abandons the traditional Ray-Ban and Oakley branding. These are purely Meta glasses now, and they are display-free. Under the hood, the device is powered by Muse Spark, making it the first consumer product to officially roll out of Meta's highly publicized superintelligence labs. Muse Spark handles real-time translation, voice-activated QA, and live scene recognition based on whatever you are looking at. It's a modest, iterative first step rather than a frontier paradigm shift, but the aggressive $299 price tag undercuts Meta's previous generation by $80. This is a cold, calculated land grab. Apple, Google, and Samsung are all racing to bring competing smart glasses to market, and Zuckerberg is intentionally nuking the entry-level pricing floor to lock in consumer distribution before his rivals can even get to the launch pad. But while Meta uses hardware to capture consumer eyes, Google Deepmind is dropping tens of millions of dollars to buy the creative respect of Hollywood's elite. Google Deepmind just cut a $75 million check to A24, the fiercely independent studio behind cinematic masterpieces like Everything Everywhere All at Once, Mid-Somar, and Heredity. A24 is arguably the most creatively respected voice in modern filmmaking, which makes this alliance a massive psychological pivot. DeepMind's Capital is explicitly earmarked co-developed production tools across the entire filmmaking pipeline, from script drafting to post-production editing, aiming to give smaller indie budgets the visual scale of a major studio. We've spent months tracking the entertainment industry's bitter legal warfare against AI, with record labels suing over training data and streaming platforms deploying forensic scanners to demonetize synthetic tracks. A24 taking Google's money completely flips that adversarial posture. This is Hollywood's creative vanguard essentially deciding that instead of fighting the technology in a courtroom, they want a seat at the table to dictate how the tools are actually built. Whether the broader creative community follows their lead or brands this as a corporate defection remains to be seen. But if anyone has the cultural capital to bring AI into the cinema world without alienating artists, it's A24. And that's it. If you have any feedback about this show, you can email Mike at yesterdaynaai.news, or you can find me on LinkedIn, X or Blue Sky. And if you like this podcast and want to see it continue, please take a moment to rate and review it so others can find it. Thanks. That's all for this edition of Yesterday and AI. Stay curious, and I'll see you tomorrow.