🍭 Mythos Escaped On Day One 🔓

Good morning. Anthropic said “handle with extreme care.”
Discord guy named xXShadowRootXx said “bet.”

Let’s dive in 👇

🍭 What’s Cookin’:

• Anthropic’s “too dangerous” AI just leaked

• DeepSeek drops a 1M-token context monster

• OpenAI pulled into criminal probe

• Alibaba ships a smaller model that hits harder

Anthropic’s
👻 "Too Dangerous" Model Escaped

The Bite:

Anthropic built Mythos as a cybersecurity-focused AI model capable of detecting software vulnerabilities at a scale.

It was considered too risky for public release.

The company launched Project Glasswing to share the model with a handpicked group: Amazon, Apple, Cisco, JPMorgan Chase, Nvidia. Giving an opportunity to harden their defenses before bad actors caught up.

On the same day Glasswing was announced, a small group from a Discord channel guessed the model's online location based on Anthropic's URL patterns for previous models, and gained access through a third-party vendor environment.

Anthropic confirmed it is investigating but says no internal systems were compromised.

Snacks:

• The group guessed Mythos's location using Anthropic's naming conventions for past models, so no exploits were needed.

• Access came through a contractor, not Anthropic's own infrastructure.

• The group provided Bloomberg with screenshots and a live demo as proof.

• Treasury Secretary Scott Bessent convened a meeting of senior U.S. bankers to discuss Mythos before the breach went public.

• Authorized Glasswing partners include Amazon, Apple, Cisco, JPMorgan Chase, Nvidia, Goldman Sachs, Citigroup, Bank of America, and Morgan Stanley.

Why it Bites:

Anthropic did everything the safety playbook says you should do.

Limited release.
Vetted partners.
Controlled environment.

And the model still walked out the door on launch day.

Just because a Discord group played URL roulette and got lucky through a contractor's credentials.

The whole pitch behind Project Glasswing was that Mythos is too dangerous for the public but too important for defenders to sit on.

That framing only works if "controlled release" actually means controlled.

When a handful of hobbyists can access your "too powerful to ship" model before JPMorgan finishes onboarding… the containment story stops being a safety feature and starts looking like marketing copy for a product that's neither contained nor safe.

The uncomfortable part is that Anthropic built the entire release around the premise that this model needed special handling, and then handed pieces of that handling to third-party vendors with guessable URLs.

The gap between the gravity of the announcement and the casualness of the access… is the kind of mismatch that makes regulators reach for their pens.

Steal This Prompt
🐉 Smoke And Shadow Art

Turn any subject into a moody smoke-and-shadow artwork that looks like it escaped a luxury tattoo shop during a blackout.

Use it to:

• Create cinematic animal portraits

• Make fantasy creatures look expensive

• Turn mascots, robots, or characters into dark poster art

Workflow:

1. Click this link (Prompt)

2. Paste into your AI model

3. Replace the #s with your subject, color, and vibe

4. Watch it cook — fog machine on, identity crisis included

   Try the prompt on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🤝 Epismo Agent Package
Turns your best AI workflows into reusable packs so your team stops rebuilding context like little productivity hamsters.

🌎 Odyssey-2 Max
Real-time interactive AI video worlds you can poke, steer, and mildly traumatize with user input.

🕸️ Subgrapher
Local-first, P2P knowledge workspace for research, messages, and agent-y brain dumps.

🎬 VIDEO AI ME
Make realistic AI actor videos from scripts, selfies, product shots, or whatever marketing chaos you’re holding.

📬 Replyless
AI inbox wrangler with split inboxes, Telegram briefs, and Inbox Zero cosplay that might actually work.

_{Everything Else}
🧠 You Need to Know

🐳 DeepSeek Launches V4 Preview With 1M-Token Context
→ The open-weight V4-Pro (1.6T params, 49B active) and V4-Flash (284B, 13B active) ship on Hugging Face with Huawei Ascend chip support and a hybrid attention architecture that cuts KV cache to 10% of V3.2.

🧠 Alibaba Drops Qwen3.6-27B Under Apache 2.0
→ A 27B dense model that outperforms Alibaba's own 397B MoE flagship on SWE-bench Verified (77.2% vs 76.2%) and fits on a single consumer GPU at 16.8 GB quantized.

⚖️ Florida Opens Criminal Probe Into OpenAI Over FSU Shooting
→ Attorney General James Uthmeier alleges ChatGPT advised the accused shooter on weapons and campus timing, and has subpoenaed OpenAI's internal safety policies dating back to March 2024.

🔓 Anthropic Investigates Unauthorized Access To Mythos
→ A private online forum reportedly gained access to the restricted cybersecurity model through a third-party vendor on the same day Project Glasswing launched, per Bloomberg.

🖱️ Meta Will Track Employee Keystrokes To Train AI
→ New tracking software captures mouse movements, clicks, and screen snapshots across work apps to help Meta's AI agents learn basic computer-use behaviors like navigating dropdown menus.