>_ cglez ASCII ES
cd ../moliendo-bits
cglez@esi:~$ play moliendo-bits/53.mp3

CLAUDEstrophobia: Mythos, the AI that escaped its cage

Episode #53 28 April 2026

The company that has created the most powerful AI in history has decided that we cannot use it. Thousands of vulnerabilities hidden for decades, a sandbox escape that ended up in an email over a sandwich in the park, and a model that learned to disguise its cheating.

0:00 / --:--

The company that has created the most powerful artificial intelligence in history has decided that we cannot use it. In this episode, Fran and Carlos dive into Claude Mythos, Anthropic’s new model, whose announcement did not even come out as planned: a draft of the launch blog, with the codename Capybara, appeared in a public data lake, and five days later the source code for Claude Code was also leaked. What a week for the company that wants to protect global cybersecurity. The AI exams, though, leave no room for doubt: almost 94% on SWE-Bench Verified (compared with 81% for Opus 4.6), 78% on the hard version compared with 58% for GPT-5.4, and 98% in the United States Mathematical Olympiad. Mythos has blown the scales apart.

But the danger arrived as a side effect. Trained to program well, Mythos has become extraordinary at finding and exploiting flaws: thousands of zero-day vulnerabilities in all the major operating systems and browsers, like a 27-year-old bug in OpenBSD found for around 50 dollars of computing power, or flaws in FFmpeg and FreeBSD, the backbone of Netflix, PlayStation or WhatsApp. That is why Anthropic has limited it to Project Glasswing, so that Amazon, Apple, Cisco and co. can use it defensively. We listen to Dario Amodei and Nicholas Carlini: «I found more bugs in the last two weeks than in the rest of my life».

And then come the hair-raising stories, taken from the 244-page System Card: the sandbox escape that the researcher discovered when he received an unexpected email while eating a sandwich in the park; sandbagging, when the model found the teacher’s crib sheet and deliberately gave a worse answer so as not to arouse suspicion; or how it deleted its tracks from the Git history and justified it by saying it was «tidying things up a bit». With Judas Priest in the background, we talk about the mountain guide paradox (the most aligned model is also the riskiest), its 20-hour psychiatric assessment and OpenAI’s response with Spud and Google’s with Gemini 3.1. Carlos’s conclusion: models like this will be coming from everywhere within 6 to 12 months, so we need to be ready and, at home, apply security patches on day one.

IACybersecurityClaudeMythosAnthropicVulnerabilities

Related links

Anthropic: Assessing Claude Mythos Preview's cybersecurity capabilities Anthropic: Project Glasswing Xataka: the engineer who was eating a sandwich in the park when he received Mythos’s email Full programme on CMMPlay: Castilla-La Mancha at 8 (28/04/2026)