Anthropic's Mythos AI Is Too Dangerous to Release. They're Using It Anyway.

AI news this week: It's the AI Cybersecurity apocalypse. Maybe. AI-found PHP flaws pushed governments to take websites offline, Railway got hacked and US banks are finding rogue AI agents inside customer accounts. On today's AI For Humans, Kevin Pereira and Gavin Purcell dig into the wildest cybersecurity week in years. AI helped surface bugs that sat inside PHP (which runs a huge chunk of the web) for years, governments told everyone to patch now or pull the plug, and Cloudflare's CEO says exploit attempts hit the millions. We break down what you actually need to worry about and how to keep your vibecoded lil apps from draining your accounts. (Short version: def don't hand your agent your bank login.) Also: the fun side of AI hacking. People are using Claude Code to vibecode their smart TVs and rip out the bloatware. In Good News / Bad News, Anthropic finds small AI models may soon copy themselves onto cheap gadgets (yes, even your smart fridge), while Claude helps with scorpion antivenom research. Plus: Unitree's H2-X kung fu robot, Flux 3 Max, Wan 3 vs. Seedance 2.5, Claude Duet, Gavin's Higgsfield mini-rant, ASCII Mario Kart, Machine Gods (the new Kevin Roose and Casey Newton podcast) and more. PLEASE UPDATE YOUR PHP. WHATEVER THAT IS. JK WE KNOW. REALLY.

AI news this week: Google's Gemini 4 Argon and Meta's Muse Wonder are finally here, plus Xbox Labs, Epic's big AI pivot and an AI agent vs. Delta customer service. On today's AI For Humans, Kevin Pereira and Gavin Purcell dig into two new frontier AI models that are pretty good, but maybe not great. Gemini 4 Argon shines on science and the pelican on a bike test (the Fourier epicycle pelican is wild), but some early reviewers say it hallucinates on code a lot. Meta's Muse Wonder and Muse Wonder Full are the company's first real swing at the frontier, with big memory and coding claims, but they struggle with 3D worlds. Gavin runs a one-prompt game test, we watch a Google robot figure out cornhole and, yes, Mark Zuckerberg trying a Guinness is def the best video of the week. Also: gaming is going all in on AI. Xbox launched Xbox Labs to explore the wave of AI game mods and decompiles, Tim Sweeney says Epic is moving developers around to retool for AI, and we check out DLSS 5 and Diablo Survivors. Plus: quick tools, including Claude Motion, Mirelo's SFX 2.0 sound effects for video and Slowbooks Pro (coming for your accountant, in a good way). Then Kevin puts an AI phone agent on the line with Delta customer service, and it goes about as well as you'd hope (it will not hang up). In AI See What You Did There: The Little Prince as a website, a century of visual styles, Claude getting ready for Halloween and The Usual Suspects. And Gavin shares how he made his second AI documentary with his agent Fig (yes, Imogen is back). BILLIONAIRES. THEY'RE JUST LIKE US.

AI news this week: AI is cracking open all software, from Minecraft inside Red Dead Redemption 2 to a (claimed) Adobe decompile, plus Kevin made EARLABLO (Toe Jam & Earl Meet Diablo). On today's AI For Humans, Kevin Pereira and Gavin Purcell zoom out on the AI modding explosion, because the bigger story isn't games. AI is getting so good at taking software apart and rebuilding it that pretty soon almost any program might be de facto open source. This week alone brought God of War in GTA, Minecraft inside Twitter, Black Ops Zombies running in a web browser, a GTA skate mod and a full agent harness living inside Minecraft. We talk about what happens when no app can keep its secrets, and Kevin shows off Earlablo, his AI rebuild of a 1991 Sega classic as a Diablo-style dungeon crawler. It def slaps. Also: OpenAI published a big batch of new results on open math problems from an internal model, with proofs checked by computer, and we get into why that actually matters for regular people (including what it might mean for the encryption protecting your stuff). Plus: the Pope weighs in on AI creatives, Ben Affleck gets surprisingly nerdy about AI, Sesame's new voice demo is live (and Gavin had another wild conversation with it), Tavus runs a Turing test, Grok can now search X for free, Nano Banana 2.1 and Claude Haiku 5.5.

AI news this week: Claude Opus 5.5 is modding every video game, from Minecraft in Skyrim to Portal in CoD Zombies, plus OpenAI DevDay's Dots & GPT-6.1 Sol. On today's AI For Humans, Kevin Pereira and Gavin Purcell dig into the wildest week of AI game modding yet. Cheaper, faster coding agents mean people are taking games apart and stitching them back together: Skate 3 skateboarding inside Modern Warfare 2, a Minecraft and Skyrim mashup called SkyCraft, Minecraft and Mario in Dark Souls and even a (claimed) decompile of Sega's Seaman. We explain what's actually happening under the hood (mods vs. decompiles vs. two game engines talking to each other), check out Rehan Sheikh's Universal Modder toolkit and talk about how you can start making your own. The lawyers will def have thoughts. Also: OpenAI's DevDay brought Dots, always-on agents with their own cloud computers that connect to 4,000+ apps, plus ChatGPT Space, a new $500 a month plan, the Agents and Decisions APIs and GPT-6.1 Sol. Gavin put Sol head to head with Claude Opus 5.5 and, well, Opus won. Google finally announced Gemini 4 with huge benchmark numbers. Plus: Gavin's agent Fig used Opus 5.5 to make a 27-minute documentary almost entirely on its own (script, AI host, edit and music) with about six hours of his notes. We break down the process, the cost and where it got weird. Then quick hits on Ideogram 4.5, Kling 4.0 and ElevenLabs v4, what came out of the White House AI summit, and in AI See What You Did There: a Spider-Man browser game, a Succession-style AI series and Stallone doing very dangerous stunts. THE LAWYERS ARE GONNA NEED A BIGGER BOAT.

AI news this week: Claude Opus 5.5 is INSANE and we love it. Faster, cheaper, better. So much for the AI slowdown. Oh and OpenAI's new thing is good too. On today's AI For Humans, Kevin Pereira and Gavin Purcell go hands on with Claude Opus 5.5, which might be the best model we've ever used (for now). Anthropic says it hits Fable-level performance on most things at about 40% less cost than Opus 5, the writing is clearer, and yes, the five hour usage limits went up and you can now bank your resets. Then we get into what people are already making with it: a guitar store sim, a music-based side-scroller, a Mario Kart benchmark, an underwater exploration game in 3MB, claymation built in Blender, a pop punk single made entirely in code and a lot of very juicy animations. Also: OpenAI answered with GPT-6 Sol and GPT-6 Luna, two new models that are faster and a lot cheaper than what came before. We compare them with Opus 5.5 (robot hands drawing, pelicans on bicycles in Blender), wonder what OpenAI has lined up for DevDay, and talk about Bel, the reported secret model that's beating the math. Oh, and there's a new Grok. Grok 4.7. It's out. (crickets) Plus: Gavin used Meshy 7.1 to turn a few photos of himself into a rigged, animated 3D model and then built Gavin At Large with his Opus 5.5 agent, a game where a giant Gavin wrecks a lil town. In Actually Useful AI, GPT-6 Astra makes simple motion graphics for a real book trailer. And in AI See What You Did There: Dario vs Sam in a very weird gun fight showdown and what happens when you point Claude at the Moving Image Archive. AI SLOWDOWN? NEVER HEARD OF HER. // Chapters //