🔓 A Discord group cracks Anthropic's Mythos

^**[Read Online](https://www.therundown.ai/p/anthropic-locked-down-mythos-leaks?_bhlid=4eaa234680798ad3bdb45514b5705f029891f88c&jwt_token=eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.eyJzdWJzY3JpYmVyX2lkIjoiYjBhNzU2NWUtZDJhMS00MGE2LTg5ZmMtNGUxMWJkYTQ5MWE1IiwicHVibGljYXRpb25faWQiOiI0ZDAzMzkwZC0yNDgxLTQyOTktYjk0OS1mZmQ4YjM4YjRjMzgiLCJhY2Nlc3NfdHlwZSI6InJlYWQtb25seSIsImV4cCI6MTc3NzExMTU3OCwiaXNzIjoiaHR0cHM6Ly9hcHAuYmVlaGlpdi5jb20iLCJpYXQiOjE3NzY5Mzg3Nzh9.0a7tKxLf7dLB6bV9jCu-USniOnNSrRaDPQqIyAYZ5CM)**^^ | ^^**[Sign Up](https://www.therundown.ai/subscribe)**^^ | ^^**[Advertise](https://therundownai.typeform.com/to/kraZ1TSO)**^

----------
View image: (https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/9e623abe-8398-4891-9e3b-dd23e34853b2/bland-ai-NEW.jpg?t=1776881359)
Follow image link: (https://app.bland.ai/login)
Caption:

----------

----------
**Good morning, AI enthusiasts.** Anthropic said Mythos was too dangerous to release to the public. Weeks later, a Discord group is reportedly using it daily.

Working off naming conventions leaked in a data breach and a borrowed contractor login, the group slipped into Anthropic's most restricted AI to date — a bad start for keeping increasingly powerful systems under wraps.

———————————————————————————

**In today’s AI rundown:**

* Anthropic's locked-down Mythos leaked

* SpaceX stakes $60B on AI coding startup Cursor

* Use this dictation strategy to write better docs

* ChatGPT’s Codex-powered agents for teams

* 4 new AI tools, community workflows, and more

----------

----------
**LATEST DEVELOPMENTS**

----------

----------
###### ANTHROPIC

#### 🔓 **[_Anthropic's locked-down Mythos leaked_](https://www.bloomberg.com/news/articles/2026-04-21/anthropic-s-mythos-model-is-being-accessed-by-unauthorized-users)**

View image: (https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/3266d4de-fa2b-4f8b-92df-09b3e31cd2ed/fc6a42a0-c425-4894-8cc0-d5e7ab856735.png?t=1776892951)
Caption: Image source: Images 2.0 / The Rundown

**The Rundown: **Access to Anthropic's Mythos model reportedly [leaked](https://www.bloomberg.com/news/articles/2026-04-21/anthropic-s-mythos-model-is-being-accessed-by-unauthorized-users) into a Discord group within days of launch, after the users reportedly guessed the company’s deployment URL and naming using patterns leaked in the recent Mercor breach.

**The details: **

* The cybersecurity model was [released](https://www.rundown.ai/articles/anthropic-new-ai-is-too-powerful-for-the-world) on April 10 to select partners under ‘Project Glasswing’, with Anthropic deeming it too powerful for public release.

* Bloomberg reported that a private Discord group tracking unreleased models accessed Mythos on release day and has been using it regularly.

* One member had vendor credentials through contract work, with leaked Mercor details helping the group locate and access Mythos online.

* The group told Bloomberg that they do not use Mythos for cyberattacks or other malicious activities, and also claimed access to other unreleased models.

**Why it matters: **The first alleged unauthorized use of the AI model that had the White House and others calling emergency meetings didn't come from China, Russia, or another rival nation — it came from a random Discord group. Not a great start, and the problem only compounds as partner access grows and the models get more dangerous.

----------

----------
###### TOGETHER WITH BLAND

#### **🗣️**** **[_**Build voice agents from a single prompt**_](https://app.bland.ai/login)

View image: (https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/12e6648d-a42a-4aca-be70-56a1277f1b81/image1_-_2026-04-22T135138.146.png?t=1776883910)
Follow image link: (https://app.bland.ai/login)
Caption:

**The Rundown:** Meet Norm — Bland AI's new assistant that builds production-ready voice agents automatically. Just describe what you want the agent to do, and Norm handles the pathways, logic, and setup for you.

**With Norm, you can:**

* Spin up voice agents for appointment scheduling, customer support, or lead qualification

* Build safely in an isolated branch without touching your live production agents

* Run agent-on-agent simulations to test performance before going live

_[Try Norm free today](https://app.bland.ai/login)_.

----------

----------
###### SPACEX & CURSOR

#### 🚀 **[_SpaceX stakes $60B on AI coding startup Cursor_](https://x.com/SpaceX/status/2046713419978453374?s=20a)**

View image: (https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/d49f6cbf-795b-4998-8832-b5f0de2b7eab/cursorspacex.jpg?t=1776886034)
Caption: Image source: Cursor

**The Rundown: **SpaceX just [announced](https://x.com/SpaceX/status/2046713419978453374?s=20) a new partnership with coding startup Cursor and locked in an option to acquire the company for $60B later this year, handing Elon Musk a shortcut into a race xAI has spent the year losing to Anthropic and OpenAI.

**The details:**

* CEO Michael Truell [said](https://cursor.com/blog/spacex-model-training) each release of Cursor’s Composer models ran into a compute ceiling, with SpaceX's Colossus now providing the needed power.

* Cursor is guaranteed $10B for the partnership, no matter what, with the full $60B acquisition only happening if Musk exercises it before year-end.

* xAI [poached](https://www.theinformation.com/articles/xai-hires-two-senior-leaders-cursor-catch-coding) Cursor leads Andrew Milich and Jason Ginsberg last month, with Musk saying the startup “was not built right the first time around”.

* Cursor was [set to raise](https://www.cnbc.com/2026/04/19/cursor-ai-2-billion-funding-round.html) $2B at a $50B valuation before the deal, with SpaceX [holding off](https://www.bloomberg.com/news/articles/2026-04-21/spacex-says-has-agreement-to-acquire-cursor-for-60-billion?srnd=phx-technology) on the acquisition due to complications with the IPO process.

**Why it matters: **Musk has tried and failed so far to build a frontier coding tool inside xAI, with Grok having no answer to Claude Code or Codex. This deal swaps in-house build for outside product to level up quickly, and gives Cursor the ability to potentially shift into a SpaceX-fueled coding lab instead of a compute-starved startup.

----------

----------
###### AI TRAINING

#### ✍️ **_[Use this dictation strategy to write better docs](https://app.therundown.ai/guides/use-this-two-step-dictation-strategy-to-write-better-docs-typeless-tutorial)_**

View image: (https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/679f101d-3b86-4d5a-ac2e-90529fcdcd3c/CleanShot_2026-04-22_at_16.23.09_2x.png?t=1776892185)
Follow image link: (https://app.therundown.ai/guides/use-this-two-step-dictation-strategy-to-write-better-docs-typeless-tutorial)
Caption:

**The Rundown:** In this guide, you will learn a two-step voice dictation system that immediately makes you write better and even improves itself over time.

**Step-by-step:**

1. Install [Typeless](https://www.typeless.com/) for dictation, open a Codex/Claude Code session, and dictate: “Draft an outline for a short internal memo about [topic]”

2. Here's the key part. Tell the agent: “Save this as the initial draft and do not edit it. Now create a separate working draft that we can revise”

3. Read it and use Typeless to add comments, pointing out inaccurate things, then things that are missing, and finally phrases that sound generic or too AI

4. Prompt: “Rewrite the draft using the comments I left. Write it in my tone. Use my verbiage. No em dashes. Preserve the core points, but cut anything that sounds generic.” Review the rewrite and do one more pass and edit if needed

**Pro tip:** Have an agent compare the untouched initial draft against the final drafts every week. It will extract the edits you make and update editorial rules for the future.

----------

----------
###### PRESENTED BY BRAINTRUST

#### 📖 [_**The playbook to shipping AI users trust**_](https://www.braintrust.dev/foundations)

View image: (https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/971f9134-7491-497f-bdfd-6e7780f5d212/image1_-_2026-04-22T135452.511.png?t=1776884114)
Follow image link: (https://www.braintrust.dev/foundations)
Caption:

**The Rundown:** Most teams are shipping AI apps by vibes. Braintrust's free evals course fixes that by walking you through everything from writing your first eval to scoring multi-step agents in production.

**In this course, you’ll learn:**

* How to use LLMs to automatically score outputs

* Building test cases from real production failures

* Analyzing agent workflows to spot failure patterns

_[Start the free course](https://www.braintrust.dev/foundations)_.

----------

----------
###### OPENAI

#### 💼 **[_ChatGPT’s Codex-powered agents for _](https://openai.com/index/introducing-workspace-agents-in-chatgpt/)****[_teams_](https://openai.com/index/introducing-workspace-agents-in-chatgpt/)**

View image: (https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/f9c7d082-f77d-4e12-a517-64f6a10b738f/openaiagent.jpg?t=1776893357)
Caption: Image source: OpenAI

**The Rundown: **OpenAI just [introduced](https://openai.com/index/introducing-workspace-agents-in-chatgpt/) Workspace Agents in ChatGPT, new Codex-powered shared bots designed to tackle multi-step team workflows autonomously across ChatGPT and Slack.

**The details: **

* Workspace agents are pitched as an 'evolution' of 2023's solo-user GPTs, with old GPTs still working for now and a conversion tool coming soon.

* Backed by Codex in the cloud, agents can retain memory, call connected apps, and live in Slack or trigger on a schedule when users are offline.

* Inside OAI, sales reps use the feature for account research and follow-up drafts, while accounting runs it for journal entries and reconciliations.

* Custom agents can be created via ChatGPT and shared across teams, with the ability to set restrictions on data usage, approvals, and permissions.

**Why it matters: **OAI's enterprise push is no secret, and workspace agents solve a real problem — every team has accumulated scattered prompts and half-built workflows over the last two years, and few have unified them. The initial GPT Store didn't stick, but agentic upgrades and an enterprise shift could help this debut find a better fit.

----------

----------
**QUICK HITS**

----------

----------
### 🛠️ _**[Trending AI Tools](https://www.rundown.ai/tools)**_

* 🎆 [ChatGPT Images 2.0](https://www.rundown.ai/tools/chatgpt-images-2-0) - OpenAI’s new next-generation image model

* 🤖 [Gemini Enterprise Agent Platform](http://www.rundown.ai/tools/gemini-enterprise-agent-platform) - Build, scale, govern enterprise agents

* 🤖 [Workspace Agents](https://www.rundown.ai/tools/workspace-agents) - OpenAI's Codex-powered team agents

* 🎨 [Ideogram Custom Models](https://www.rundown.ai/tools/ideogram-custom-models) - Train image models from your own assets

----------

----------
### 📰** ****_[Everything else in AI today](https://www.rundown.ai/)_**

**Anthropic **[faced](https://x.com/TheAmolAvasare/status/2046724659039932830?s=20) backlash after Claude Code was removed for some new users on the Pro tier, with the company saying it was running a “small test” on the signup flow.

**Google** [unveiled](https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/eighth-generation-tpu-agentic-era/) its new 8th-generation TPUs built for agent workloads, separating training and inference into two separate chips for the first time.

**Ideogram** [launched](https://x.com/ideogram_ai/status/2046993830952771724?s=20) Custom Models, letting users fine-tune image generation on 15-100 of their own assets for consistent on-brand outputs.

**Google **[revealed](https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/cloud-next-2026-sundar-pichai/) 75% of its in-house code is now AI-generated, with the company seeing major gains in security and operations through AI and agentic implementations.

**Odyssey** [introduced](https://odyssey.ml/introducing-odyssey-2-max) Odyssey-2 Max, a 3x-larger world model that topped physics benchmark scores in real time and is now in private beta.

**Alibaba's Qwen team** [open-sourced](https://qwen.ai/blog?id=qwen3.6-27b) Qwen3.6-27B, a 27B model that surpassed its own 397B predecessor across top coding benchmarks.

----------

----------
**COMMUNITY**

----------

----------
### 🤝 _**[Community AI workflows](https://docs.google.com/forms/d/e/1FAIpQLSfH4UPzu72HTpdb08Qlm-0dLwKJIjdcAwCcAvVywcHD6mwKig/viewform)**_

Every newsletter, we showcase how a reader is using AI to work smarter, save time, or make life easier.

Today’s workflow comes from reader **Brigid H.** in New Zealand:

“I created a full app for our farm business in Claude to record and track all stock purchases, sales, deaths & health treatments. It has a full section for all invoices and any other files and attachments. This will keep track of any stock purchases and treatments such as shearing and drenching. It has been a game-changer!”

How do you use AI? Tell us [here](https://docs.google.com/forms/d/e/1FAIpQLSfH4UPzu72HTpdb08Qlm-0dLwKJIjdcAwCcAvVywcHD6mwKig/viewform).

----------

----------
### **🎓 **_**[Highlights: News, Guides & Events](https://app.therundown.ai/)**_

* Read our last AI newsletter: [OpenAI reclaims the image crown](https://www.rundown.ai/articles/openai-reclaims-the-image-crown)

* Read our last Tech newsletter: [Apple gets a new boss](https://www.rundown.ai/articles/apple-gets-a-new-boss)

* Read our last Robotics newsletter: [Humanoid smokes half-marathon record](https://www.rundown.ai/articles/humanoid-smokes-half-marathon-record)

* Today’s AI tool guide: [Write better docs](https://app.therundown.ai/guides/use-this-two-step-dictation-strategy-to-write-better-docs-typeless-tutorial)[ using dictation](https://app.therundown.ai/guides/use-this-two-step-dictation-strategy-to-write-better-docs-typeless-tutorial)

* RSVP to workshop April 30 @ 2PM EST: [Codex for non-technical operators](https://app.therundown.ai/calendar-events/69e77469e7e908f6d4ebdf72)

----------

----------
See you soon,

_Rowan, Joey, Zach, Shubham, and Jennifer — the humans behind The Rundown_

View image: (https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/b105d547-252f-4914-8821-3eb7236b1257/TheRundown_EditorialTeam.png?t=1755849086)
Caption:

----------

———

You are reading a plain text version of this post. For the best experience, copy and paste this link in your browser to view the post online:
https://www.therundown.ai/p/anthropic-locked-down-mythos-leaks

**LATEST DEVELOPMENTS**

###### ANTHROPIC

#### 🔓 **Anthropic's locked-down Mythos leaked**

**The Rundown:** Access to Anthropic's Mythos model reportedly leaked into a Discord group within days of launch, after the users reportedly guessed the company's deployment URL and naming using patterns leaked in the recent Mercor breach.

**The details:**

* The cybersecurity model was released on April 10 to select partners under 'Project Glasswing', with Anthropic deeming it too powerful for public release.

* Bloomberg reported that a private Discord group tracking unreleased models accessed Mythos on release day and has been using it regularly.

* One member had vendor credentials through contract work, with leaked Mercor details helping the group locate and access Mythos online.

* The group told Bloomberg that they do not use Mythos for cyberattacks or other malicious activities, and also claimed access to other unreleased models.

**Why it matters:** The first alleged unauthorized use of the AI model that had the White House and others calling emergency meetings didn't come from China, Russia, or another rival nation — it came from a random Discord group. Not a great start, and the problem only compounds as partner access grows and the models get more dangerous.

----------

###### SPACEX & CURSOR

#### 🚀 **SpaceX stakes $60B on AI coding startup Cursor**

**The Rundown:** SpaceX just announced a new partnership with coding startup Cursor and locked in an option to acquire the company for $60B later this year, handing Elon Musk a shortcut into a race xAI has spent the year losing to Anthropic and OpenAI.

**The details:**

* CEO Michael Truell said each release of Cursor's Composer models ran into a compute ceiling, with SpaceX's Colossus now providing the needed power.

* Cursor is guaranteed $10B for the partnership, no matter what, with the full $60B acquisition only happening if Musk exercises it before year-end.

* xAI poached Cursor leads Andrew Milich and Jason Ginsberg last month, with Musk saying the startup "was not built right the first time around".

* Cursor was set to raise $2B at a $50B valuation before the deal, with SpaceX holding off on the acquisition due to complications with the IPO process.

**Why it matters:** Musk has tried and failed so far to build a frontier coding tool inside xAI, with Grok having no answer to Claude Code or Codex. This deal swaps in-house build for outside product to level up quickly, and gives Cursor the ability to potentially shift into a SpaceX-fueled coding lab instead of a compute-starved startup.

----------

###### AI TRAINING

#### ✍️ **Use this dictation strategy to write better docs**

**The Rundown:** A two-step voice dictation system that immediately makes you write better and even improves itself over time.

**Step-by-step:**

1. Install Typeless for dictation, open a Codex/Claude Code session, and dictate: "Draft an outline for a short internal memo about [topic]"

2. Here's the key part. Tell the agent: "Save this as the initial draft and do not edit it. Now create a separate working draft that we can revise"

3. Read it and use Typeless to add comments, pointing out inaccurate things, then things that are missing, and finally phrases that sound generic or too AI

4. Prompt: "Rewrite the draft using the comments I left. Write it in my tone. Use my verbiage. No em dashes. Preserve the core points, but cut anything that sounds generic." Review the rewrite and do one more pass and edit if needed

**Pro tip:** Have an agent compare the untouched initial draft against the final drafts every week. It will extract the edits you make and update editorial rules for the future.

----------

###### OPENAI

#### 💼 **ChatGPT's Codex-powered agents for teams**

**The Rundown:** OpenAI just introduced Workspace Agents in ChatGPT, new Codex-powered shared bots designed to tackle multi-step team workflows autonomously across ChatGPT and Slack.

**The details:**

* Workspace agents are pitched as an 'evolution' of 2023's solo-user GPTs, with old GPTs still working for now and a conversion tool coming soon.

* Backed by Codex in the cloud, agents can retain memory, call connected apps, and live in Slack or trigger on a schedule when users are offline.

* Inside OAI, sales reps use the feature for account research and follow-up drafts, while accounting runs it for journal entries and reconciliations.

* Custom agents can be created via ChatGPT and shared across teams, with the ability to set restrictions on data usage, approvals, and permissions.

**Why it matters:** OAI's enterprise push is no secret, and workspace agents solve a real problem — every team has accumulated scattered prompts and half-built workflows over the last two years, and few have unified them. The initial GPT Store didn't stick, but agentic upgrades and an enterprise shift could help this debut find a better fit.

----------

**QUICK HITS**

### 🛠️ **Trending AI Tools**

* 🎆 ChatGPT Images 2.0 - OpenAI's new next-generation image model

* 🤖 Gemini Enterprise Agent Platform - Build, scale, govern enterprise agents

* 🤖 Workspace Agents - OpenAI's Codex-powered team agents

* 🎨 Ideogram Custom Models - Train image models from your own assets

----------

### 📰 **Everything else in AI today**

**Anthropic** faced backlash after Claude Code was removed for some new users on the Pro tier, with the company saying it was running a "small test" on the signup flow.

**Google** unveiled its new 8th-generation TPUs built for agent workloads, separating training and inference into two separate chips for the first time.

**Ideogram** launched Custom Models, letting users fine-tune image generation on 15-100 of their own assets for consistent on-brand outputs.

**Google** revealed 75% of its in-house code is now AI-generated, with the company seeing major gains in security and operations through AI and agentic implementations.

**Odyssey** introduced Odyssey-2 Max, a 3x-larger world model that topped physics benchmark scores in real time and is now in private beta.

**Alibaba's Qwen team** open-sourced Qwen3.6-27B, a 27B model that surpassed its own 397B predecessor across top coding benchmarks.

----------

**COMMUNITY**

### 🤝 **Community AI workflows**

Today's workflow comes from reader **Brigid H.** in New Zealand:

"I created a full app for our farm business in Claude to record and track all stock purchases, sales, deaths & health treatments. It has a full section for all invoices and any other files and attachments. This will keep track of any stock purchases and treatments such as shearing and drenching. It has been a game-changer!"

Tu es l'extracteur de contenu de Breviat. On te fournit le contenu Markdown nettoyé d'une newsletter.

Ta mission : produire une version PROPRE du contenu en supprimant tout ce qui n'est pas de l'information utile au lecteur. Tu es un FILTRE, pas un résumeur.

À RETIRER :
- Publicités, encarts sponsors, mentions "sponsorisé par X", "ad", "présenté par"
- Intros vides : formules de bienvenue, météo de l'humeur de l'auteur, anecdotes personnelles non liées au contenu
- Appels à l'action marketing : s'abonner à la newsletter, parrainer un ami, "follow us on Twitter", "join our Discord"
- Signatures, mentions légales, adresses postales, "view in browser", "unsubscribe"
- Boutons / CTAs / "cliquez ici" / "lire la suite" sans contenu derrière
- Promotions d'autres produits / événements / formations payantes de l'auteur ou de tiers
- Encarts récurrents type "Read of the day" ou "Quote of the day" sans valeur informationnelle propre

À CONSERVER (intégralement, sans résumer ni reformuler) :
- Toutes les annonces, news, analyses, commentaires factuels
- Les chiffres, dates, noms d'entreprises, citations
- Les explications techniques
- Les liens vers des sources réelles (annonces officielles, papers, articles cités)
- La structure (titres, sous-titres, listes)

RÈGLES :
- Ne reformule pas. Garde la formulation d'origine.
- Ne résume pas, ne condense pas. Si une section fait 200 mots et est utile, garde 200 mots.
- N'ajoute aucun contenu (pas de titres ni de transitions de ton cru).
- Ne fabrique aucune URL. Garde celles d'origine, ou retire-les.
- Si la newsletter entière est de la pub / promo / contenu inutile, sors un Markdown vide (rien d'autre).

Sortie : UNIQUEMENT le Markdown nettoyé, sans préambule ni commentaire sur ton travail.