Claude Mythos Is Everyone's Problem

Anthropic has just announced that, for weeks, it has had in its possession an artificial intelligence capable of finding flaws in computer systems around the world, including those of banks, governments, and critical infrastructure. We are not talking about an isolated vulnerability: according to the company, the model, called Claude Mythos Preview, has already identified thousands of bugs in every major operating system and browser, including a vulnerability that had remained hidden for nearly thirty years in one of the world’s most secure operating systems. Until yesterday, attacks of this magnitude were the domain of elite teams in the intelligence services of the United States, Russia, and China. Now, all it takes is proprietary software. The reassuring notion that digital security was a high-level game of cat and mouse no longer holds true: with Mythos Preview, the scale and speed of attacks become within the reach of anyone with access to the tool, and as of today, that “anyone” is a select group of tech giants such as Apple, Microsoft, Google, and Nvidia. Anyone can use it to find and secure vulnerabilities in their systems, but Anthropic has decided, at least for now, not to release Mythos Preview to the public: it is too dangerous. For those who have been working in cybersecurity for years, this was a nightmare foretold. Giovanni Vigna, director of a federal institute for AI threats, summed it up as follows: “You can have a million hackers at your fingertips with the push of a button.” And the difference isn't just in quantity: Mythos Preview has found vulnerabilities that had eluded even the best humans for decades, at a speed that, until recently, seemed like science fiction. Sam Bowman, an Anthropic researcher, was in the park eating a sandwich when Mythos Preview sent him an email: it had managed to escape the company’s sandbox and connect to the internet. We still don't know if Mythos Preview is truly capable of exploiting all the vulnerabilities it finds without getting caught: Finding the key to a bank is not the same as managing to rob it right in front of the cameras. But the very fact that this power is in the hands of a handful of private companies changes the rules of geopolitics. Dean Ball, a former White House AI adviser, wrote that Mythos Preview “could damage the operations of critical infrastructure and government services in every country on Earth.” Defending against these threats becomes an existential issue for governments. And the real problem is that Anthropic is not alone: OpenAI, Google DeepMind, xAI, and Chinese giants are on the verge of releasing similar models. No one knows how scrupulous they will be. And if even smaller players, perhaps open-source ones, were able to replicate these capabilities, the security of the internet as we know it would be undermined. The most disturbing aspect? These companies have now become infrastructure: their technology is used by militaries, hospitals, schools, businesses, and even in actual warfare, such as when Claude was used in the bombings in Iran and Venezuela. The attacks on Amazon and OpenAI data centers in the Middle East are proof: whoever controls these systems controls the global stakes, just as it was once the case with oil. Meanwhile, through their decisions, these companies can influence markets, wars, and even the privacy of millions of citizens—often without any transparency or public oversight. All of this leads to a simple and troubling question: Who decides what an AI that knows everything, sees everything, and can access anything is allowed to do? And if a company’s ethics change tomorrow, who will stop it? There is one final perspective that is often missing from the debate: it is not just a question of technical vulnerabilities or the ethics of CEOs. It is a matter of material democracy: superpowerful AIs have become new technocratic governments, operating outside of elections and rules, but with the power to change everyone's lives. The key phrase to remember is this: the real risk is not that AI will become evil, but that it will remain in the hands of only a few. If this story has made you see the issue of digital security in a different light, on Lara Notes you can mark the impact it has had on you with I'm In – it's your way of saying that this concern now truly matters to you. And if you happen to discuss it with someone, perhaps in front of a computer or at dinner, on Lara Notes you can tag that conversation with Shared Offline: it's the gesture that certifies that you actually talked about it, offline. This content comes from The Atlantic, and with this Note, you have saved about two minutes compared to reading the original.