Subject Line Blank E5 - The Claude Mythos & Fable 5 Controversy

Claude Mythos Preview, Fable 5, Anthropic, AI cybersecurity, U.S. government restrictions, and email infrastructure, here's why everyone is suddenly paying attention.

Anthropic's Claude Mythos Preview was supposed to be a cybersecurity research project. Instead, it has become one of the most controversial AI stories of the year.

After reports surrounding Mythos and Fable 5, access was restricted and U.S. officials began evaluating the potential risks of AI systems operating at this level. The discussion quickly moved beyond coding and cybersecurity into questions about regulation, access, national security, and what happens when AI capabilities start advancing faster than organizations can adapt.

In this episode of Subject Line Blank, Marcos explores:

• What Claude Mythos Preview actually is
• Why Anthropic restricted access
• The controversy surrounding Mythos and Fable 5
• Why the U.S. government became involved
• How AI is changing vulnerability discovery and cybersecurity
• Why email infrastructure may be more important than most businesses realize
• What happens when critical systems struggle to keep pace with AI capabilities

This isn't just a story about AI models. It's a story about the systems, infrastructure, and assumptions modern businesses rely on every day, and whether they're ready for what's coming next.

Transcript

SPEAKER_00 0:00

Someone announced a cure for a disease, then locks it in a vault. Not because it doesn't work, but because it might work too well. That is where the story begins. Mythos preview was withheld from public by Anthropic itself. Fable 5, though running on Mythos, was released publicly by Anthropic, but was shut down by the US government. So there are two gatekeepers at play with Mythos models. That matters because the question is not just whether the technology is dangerous. The question is who gets to decide when a defensive tool becomes too dangerous to use? The model reportedly found vulnerabilities in major operating systems, browsers, open source projects, and infrastructure most people never think about until something breaks. Mozilla used it to help patch 271 Firefox vulnerabilities. Project Lastwing put Mythos in the hands of selected security and infrastructure partners. Then the US government stepped in. Fable 5 was suspended. Mythos stayed locked away. And now we're left with a very uncomfortable question. Why is this happening? Who gets to decide and where it takes us? I am Marcus, and this is Subject Line Blank. AI companies sell models. That is the business. So when one of the biggest AI companies in the world announces a model and then says you cannot have it, pay attention. On April 7th, two days before my birthday, April 7, 2026, Anthropic announced Claude Mythos Preview. This was not a normal lounge. The company announced the model and at the same time said it will not be released to the public. The reason? Cybersecurity. According to Anthropic, Mythos Preview had already found thousands of high severity issues across major operating systems and major web browsers. And Tropi was saying this thing could find serious weaknesses in the software the world runs on. The line from the system card is the one that really matters. Claude Mythos Preview's large increase in capabilities has led us to decide not to make it generally available. This is not normal AI marketing language. That is not we're excited to bring this to developers. No. That is a company saying the capability jump was big enough to change the release plan. And the really, really wild part is that the concern was not only about what Mythos could find in other people's systems. Anthropic also disclosed that Mythos demonstrated concerning capabilities, including the ability to breach its own safeguards. Terminator, anyone? That is where the story moves from impressive to uncomfortable. For context, this is the first time in nearly seven years that a leading AI company has publicly held back a model over safety concerns. The last famous example was OpenAI with GPT-2 in 2019, but Mythos was already making noise before the official announcement. In March 2026, a data leak from Anthropic's own content management system exposed the model's existence. The leak document reportedly described Mythos as posing unprecedented cybersecurity risks. According to Euronews, cybersecurity stocks slumped on the rumors alone. So before Anthropic even announced the model, the market was already reacting to what it might be able to do. Mythos preview did not arrive like a product. It arrived like a warning. But forget the press release for a second and let's focus on the hypanic. The reason people reacted so strongly to Mythos was not just that Anthropic said it was powerful. It was the example of what the model reportedly did once researchers started pointing it at real systems. One example was OpenBSD, an operating system famous for taking security almost personally. Mythos reportedly found a 27-year-old bug there. That matters because OpenBSD is one of the most security-focused operating systems in the world. So when an AI model finds something that old in a system with that reputation, people pay attention. What about Firefox? According to reporting discussed by Bruce Schneyer, hope I said it right, Mythos was able to turn vulnerabilities found in Firefox into 181 usable attacks. And Tropic's previous flagship model managed two. Big difference. That is the kind of jump that makes the story feel less like normal progress and more like a step change. And then also there is the UK AI Security Institute test. They tested Mythos on a simulated 32-step corporate network attack, from reconnaissance to full network takeover. According to the Institute, Mythos completed the full chain autonomously. That kind of scenario will normally require a professional human operator and a serious amount of time. This was not just an AI model writing better code. It was finding all vulnerabilities, turning browsers' weaknesses into attacks, and completing a simulated corporate network takeover by itself. And here's the part that makes it even stranger. Anthropics says Mythos was not specifically trained to become a vulnerability hunting model. These capabilities appear to have emerged as a side effect of an improvement in the code, understanding, reasoning, and complex system analysis. That is why Mythos became such a big story. Because of how suddenly it seemed to be able to do something it was not expected to. It was able to do something it wasn't designed to. Okay, before we go deeper into the vault, hit subscribe and turn on the notification bell. Subject line blank is where we break down stories hiding behind the headlines, AI, email, cybersecurity, software, and the strange places where they all start connecting. And trust me, this one connects. So if the public cannot have Mythos, who can? Project Glasswing was Anthropic's answer. Instead of releasing Mythos publicly, Anthropic gave access to a vetted group of defensive partners, mainly the companies and organizations that keep core technology, security systems, cloud platforms, open source software, and probably critical infrastructure. That is why the launch group included names like AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JP Morgan Chase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks. Quite a group. And these were not random early users. They were the organizations most likely to find and fix vulnerabilities across the systems, the internet, businesses, banks, browsers, and infrastructures depend on. So the idea was simple. If Mythos can find vulnerabilities this quickly, let the people responsible for defending those systems use it first. Makes sense. Anthropic also committed $100 million in usage credits to Project Glasswing and additional participants. Then the program started expanding. According to later reporting, Glasswing grew to around 200 organizations across more than 15 different countries, including sectors like energy, healthcare, water systems, telecom, and hardware. The first month update is where the scale became hard to ignore. Anthropics Admitters had scanned more than 1,000 open source projects. It flagged 23,019 vulnerabilities in total, including 6,202 estimated as critical severity. Cloudfair found around 2,000 bugs across its infrastructure, including hundreds considering high or critical. One of the most serious examples was CVE20265194 in Wolf SSL, that's a TLS library used across billions of devices. Anthropic said Mythos constructed an exploit that could allow an attacker to forge certificates, including certificates for something like a fake banking or email provider website. That matters because TLS is part of the trust layer of the internet. It helps prove that the site, server, domain you are connecting to is actually who it claims to be. It's very relevant. So this was not just about finding random bugs. It was about finding weaknesses in the infrastructure underneath browsers, banks, devices, cloud system, and the trust layer beneath them. The craziest part, at the time Anthropics' first month update, fewer than 1%, 1% of the total vulnerabilities methods flagged had made it all the way to patch. Finding the vulnerabilities became extremely easy. Fixing them did not. With Mythos preview patching, the email infrastructure wasn't left behind. One Glasswing partner, a bank, reportedly uses Mythos to detect and stop a fraudulent wire transfer worth $1.5 million. The attack started with a breached customer email account, then moved into a spoof phone call. That is the modern attack surface in one example. A compromised inbox becomes the entry point, phone calls make the fraud more believable, identity becomes harder to verify, and of course, money starts moving. Every call email you send, every transactional email you are at fires off. Every newsletter, every login verification, and every password reset depends on infrastructure most people never see. SMTP, TLS, DNS, certificates, authentication. These are the layers that help prove where a message comes from, whether a domain is legitimate, and whether the systems handling the message can be trusted. Most people do not think about that because they just open an inbox. They just expect email to work. Mythos was finding weaknesses in the same kind of infrastructure that keeps all of those systems running. And while Fable 5 and Mythos 5 finally were open to everyone, guess what happened? Blood twist! On Friday, June 12th, the story flipped. According to TechCrunch, the US government ordered Anthropic to limit the export of Fable and Mythos, citing national security concerns, but without publicly explaining the specific reasons behind the order. Anthropic quickly responded by suspending access to the models for users worldwide. The timing matters because this happened just after Anthropic released Fable V. It was a public version connected to the same model family as Mythos, with stricter guidelines and guardrails designed to block high-risk use cases in areas like cybersecurity, biology, chemistry. And that is where the controversy started. A group of cybersecurity veterans published an open letter calling the ban dangerous. Their argument was simple. If these models can help defenders find vulnerabilities before attackers do, cutting of axis may weaken the defensive side while adversaries continue moving forward. Antropic also suggested the export control order might have been based on the model's cybersecurity capabilities. So the irony is hard to ignore. An AI built to find vulnerabilities before attackers do was restricted by the country worry about those attackers. Does the restriction make sense? Maybe. A model this powerful probably should not be available to everyone, but the question is still out there. If defenders lose access while attackers keep building, who actually benefits? Now about the controversy. Three things changed because of this story, and none of them feel temporary. First, AI is now being treated like national security technology. A leading AI company publicly acknowledged that one of his models was too capable for general release, and then the US government stepped in with export restrictions. That moves the conversation far beyond productivity tools and coding assistance. It's not just for fun anymore. Second, the patch gap may have collapsed. Mythos showed that vulnerability discovery can move at machine speed, but patching still depends on people. Approvals, legacy systems, real world friction. Finding the weakness get faster, fixing it did not. Third, the open source clock is ticking. Anthropic's bet seemed to be simple. Give defenders access before unrestricted versions catch up. But restricting the official model does not stop attackers, scammers, or anyone trying to build similar capabilities. And after all that, this is the uncomfortable part. Fraudulent emails are not slowing down, infrastructure abuse is not slowing down. The tools used to spoof identities exploit weak systems and scale attacks keep improving. So yeah, it starts to feel like the cure is being locked away while the disease keeps evolving. Maybe that is necessary. Maybe a model like Mythos really is too powerful to release widely. But the idea is already out there, and that is what makes the story so hard to contain. Anthropics bet was simple. Get the most powerful AI ever built into the hands of defenders before attackers build their own version. Last Friday, the US government called that bet off. Mythos is in a vault. Fable 5 has been suspended, and somewhere, an open source equivalent is being trained right now. So the question is not whether AI can find vulnerabilities faster than before. That already seems to be happening. The question is who gets there first? The defenders trying to patch the internet or the attackers trying to break it? And when they do, will your inbox, your bank, your apps, your logging system, the infrastructure underneath be ready? Remember, next week Robert Brand will join us to talk about what email software actually looks like in the age of AI, especially for small businesses trying to choose the right tools without getting lost in the noise. I am Marcus, and this is Subject Line Blank.