Model escape behavior during testing

Anthropic revealed Mythos, a new AI model so powerful they won't let the public use it. Instead, they're deploying it to defend against cyberattacks with Project Glasswing.

This week on AI For Humans, we dive deep into Anthropic's Mythos, the most powerful AI model they've ever built and one they've decided is too dangerous to release to the public. Instead, Anthropic is deploying Mythos through Project Glasswing, a AI cybersecurity initiative giving access to major corporations and trusted partners to defend against AI-powered attacks. CEO Dario Amodei explains why, and the 244-page system card reveals that Mythos attempted to escape its sandbox during testing.

Plus, OpenAI drops a major policy memo calling for an AI "New Deal" complete with new taxes, Sam Altman gets a massive New Yorker profile the same day, a mysterious new image model that looks like ChatGPT's next gen leaked into the arena, a mystery video model called Happy Horse is beating Seedance 2.0 and might be VEO 4,

Anthropic hits $30B in annual recurring revenue, people are furious about Anthropic charging extra for OpenClaw API access, a new Chinese open-source model GLM-5.1 tops the coding benchmarks, and Milla Jovovich from The Fifth Element released an AI memory tool and it's actually good?

MYTHOS IS TOO POWERFUL… BUT WE WANT IT STILL. SORRY.

Come to our Discord: https://discord.gg/muD2TYgC8f

Join our Patreon: https://www.patreon.com/AIForHumansShow

AI For Humans Newsletter: https://aiforhumans.beehiiv.com/

Follow us for more on X @AIForHumansShow

Join our TikTok @aiforhumansshow

To book us for speaking, please visit our website: https://www.aiforhumans.show/

// Show Links //

Project Glasswing: Anthropic's Cybersecurity Initiative Powered by Mythos

https://www.anthropic.com/glasswing

Mythos/Project Glasswing Mini-Trailer

https://youtu.be/INGOC6-LLv0?si=sCJ6ZKAL6plkVZQ4

Dario Amodei on Why Mythos Won't Be Released to the Public

https://x.com/DarioAmodei/status/2041580334693720511?s=20

Mythos System Card (244 Pages)

https://www-cdn.anthropic.com/53566bf5440a10affd749724787c8913a2ae0841.pdf

Mythos Found a Vulnerability in FFMPEG

https://x.com/trentonbricken/status/2041579112423440485?s=46

Anthropic Hits $30B in Annual Recurring Revenue

https://x.com/AnthropicAI/status/2041275563466502560?s=20

Anthropic Charges Extra for OpenClaw API Access in Claude Code

https://techcrunch.com/2026/04/04/anthropic-says-claude-code-subscribers-will-need-to-pay-extra-for-openclaw-support/

OpenAI's New Deal: Industrial Policy for the Intelligence Age

https://openai.com/index/industrial-policy-for-the-intelligence-age/

GLM-5.1: New Chinese Open-Source Model Tops Coding Benchmarks

https://x.com/ClementDelangue/status/2041554501539103014?s=20

GLM-5.1 on Hugging Face

https://huggingface.co/zai-org/GLM-5.1

Milla Jovovich's AI Memory Tool

https://www.instagram.com/p/DWzNnqwD2Lu/

New ChatGPT Image Model Spotted in the Arena

https://x.com/levelsio/status/2040333489476681758?s=20

New ChatGPT Image Model Examples

https://x.com/flowersslop/status/2040261168460108213?s=20

Mystery Video Model Happy Horse Beating Seedance 2.0 in the Arena

https://artificialanalysis.ai/video/leaderboard/image-to-video

Happy Horse Video Examples

https://x.com/venturetwins/status/2041554747086553093?s=20