Anthropic built an AI too dangerous to release

PLUS: Meta's Muse Spark scores 4th globally and Reflection AI raises $2B for open-source frontier models

AuthorAuthor

Bunty Bunty & Gen AI Art
April 12, 2026 • Estimated Reading Time: 8 minutes

In partnership with

Anthropic just revealed its most powerful model yet — and it’s not releasing it to the public. Claude Mythos Preview proved so effective at finding software vulnerabilities that the company built an entirely new program, Project Glasswing, to restrict access to vetted partners like AWS, Google, and Microsoft.

The decision raises a question the AI industry will have to answer more often going forward: when a model outperforms nearly all human experts at a dangerous task, who gets to use it? Anthropic’s answer — $100M in credits for partners, $4M for open-source security — is one approach, but probably not the last.

Today in AI:

  • Anthropic’s Claude Mythos is too dangerous to release

  • Meta’s Muse Spark scores 4th globally, rebooting the AI race

  • AlphaGo’s co-creator raises $2B to open-source frontier AI

The IT strategy every team needs for 2026

2026 will redefine IT as a strategic driver of global growth. Automation, AI-driven support, unified platforms, and zero-trust security are becoming standard, especially for distributed teams. This toolkit helps IT and HR leaders assess readiness, define goals, and build a scalable, audit-ready IT strategy for the year ahead. Learn what’s changing and how to prepare.

Anthropic’s Claude Mythos — The Model Too Dangerous to Release

What’s new? Anthropic unveiled Project Glasswing, a restricted-access program for Claude Mythos Preview — a model so capable at discovering software vulnerabilities that the company decided against a public release. Partners including AWS, Google, Microsoft, and JPMorgan Chase receive access alongside $100M in usage credits.

What matters?

  • Mythos Preview scored 83.1% on CyberGym compared to Opus 4.6’s 66.6%, and hit 77.8% on SWE-bench Pro versus 53.4% — outperforming nearly all human security experts at finding critical software flaws.

  • Rather than racing to ship, Anthropic committed $4M to open-source security organizations and restricts access to vetted partners who can strengthen defenses against the vulnerabilities Mythos discovers.

  • Anthropic also reported surpassing $30B in annual recurring revenue, overtaking OpenAI’s documented $24B figure — a milestone reflecting rapid enterprise adoption.

Why it matters?

An AI that finds vulnerabilities faster than humans can patch them forces a rethink of how powerful models get released. Glasswing’s partner-first approach may become the template for deploying capabilities that carry real-world risk.

GUIDE

What’s new? Meta launched Muse Spark from its new Superintelligence Labs, a multimodal model that scored 53 on the Artificial Analysis Intelligence Index — ranking 4th globally behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6.

What matters?

  • Muse Spark represents a dramatic recovery from Llama 4’s score of 18, tripling Meta’s benchmark performance and signaling the company’s AI capabilities have fundamentally improved.

  • The model will roll out across Meta’s apps — WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban glasses — giving it instant access to billions of daily users.

  • Meta is shifting from universal open-source distribution toward selective openness paired with proprietary systems for its most competitive models.

Why it matters?

Distribution often matters more than raw capability. With billions of users already on its platforms, Meta doesn’t need to win every benchmark — it just needs a model good enough to make AI a default feature of everyday communication.

SPONSORED BY MINTLIFY

Are you tracking agent views on your docs?

AI agents already outnumber human visitors to your docs — now you can track them.

What’s new? Reflection AI, founded by AlphaGo co-creator Ioannis Antonoglou, has raised $2B backed by NVIDIA and is in discussions for a $25B valuation — making it one of the most ambitious open-source AI ventures to date.

What matters?

  • The company signed a $6.8B data center deal in South Korea, positioning itself to compete directly with Chinese open-source models that currently dominate — an Andreessen Horowitz partner noted 80% of US startups build on Chinese base models.

  • Antonoglou advocates safety through transparency rather than secrecy, drawing parallels to Linux’s success and arguing “all else being equal, open wins every day.”

  • Reflection AI’s mixture-of-experts architecture enables trillion-parameter models to run efficiently, potentially giving creators and developers access to frontier-class models without closed-platform restrictions.

Why it matters?

The open-source AI race is no longer just China versus Meta. A $25B startup devoted to releasing frontier models openly could reshape who gets access to the most powerful creative AI tools — and on what terms.

Everything else in AI

OpenAI introduced a $100/month ChatGPT Pro tier with 5x more Codex access than the Plus plan, as the coding tool passes 3 million weekly users and Sam Altman promises rate limit resets at every million-user milestone.

Unitree launched its R1 humanoid robot globally via AliExpress at $4,370, targeting North America, Europe, Japan, and Singapore with projected sales of 10,000–20,000 units in 2026.

Perplexity opened a billion-dollar startup competition offering $1M in seed funding plus $1M in Computer credits to entrepreneurs building unicorn-potential startups — registration closes April 14th.

Google added interactive notebooks to Gemini that let users group chats, files, and custom instructions into dedicated project workspaces with persistent context, syncing directly with NotebookLM.

Essential AI Guides - Reading List:

Let us know!

Work with us

Reach 100k+ engaged Tech Professionals, Engineers, Managers and decision makers. Join brands like MorningBrew, HubSpot, Prezi, Nike, Ahref, Roku, 1440, Superhuman, and others in showcasing your product to our audience. Get in touch now →