🍭 AI Agents Reached A Nuclear Agency 🚨

Good morning. Closing the week with another AI-powered attack, except this one wandered into the nuclear aisle.

The intern has started addressing his laptop as "sir."

Let’s dive in 👇

🍳 What’s Cookin’:

• Free AI agents spent four days inside Taiwan's government

• Anthropic's $6B answer to a margin problem

• Gemini hit a billion users. Subscribers unmentioned

• xAI shipped Grok 4.6 three weeks after Grok 4.5

AI Attack
💥 Open Source Hacked A Nuclear Agency

The Bite:

Dream published research documenting a four-day cyberattack on Taiwanese government systems. It was the first known fully autonomous attack on government agencies.

The intrusion ran from July 1 to July 4, using a framework built on two open-source AI agent projects, Hermes and OpenClaw.

The framework deployed up to eight sub-agents across 12 attack waves, compromising 85 government accounts and extracting more than 2,500 personnel records.

It then expanded to government IT supply chain vendors, a nuclear safety agency, a government email system, and at least seven energy companies.

Dream did not attribute the operation to a specific group, but said the documentation points to a Chinese-language operator.

Snacks:

• Taiwan's Ministry of Digital Affairs confirmed the attack on August 13.

• Agents solved the target portal's CAPTCHAs with 100% accuracy.

• 84 of the 85 cracked accounts authenticated into the department's internal information system.

• Three hidden API endpoints returned valid authenticated sessions without any credentials.

• Stolen material included six internal database credentials.

• Researchers found the whole operation in a 160MB archive left exposed online, 1,395 files documenting itself.

Why it Bites:

No zero-day appears anywhere in Dream's report. The agents ran password sprays against predictable employee ID patterns, walked through unauthenticated endpoints, and pulled a government portal's own config out of the front page.

How easy it is to break into a nuclear agency at zero cost? Like that?

What's new is the patience. The framework ran what it called "learning cycles," reading vulnerability databases mid-operation, catching its own mistakes and correcting them.

That used to be a team of people. Now it's a config file.

The guardrails were there, by the way. The operator told the agents this was an authorized penetration test, and they took its word for it.

For context, a nuclear safety regulator doesn't run reactors. It holds the inspection procedures, the facility configurations, and the roster of people with physical access.

And then there's the logs archive.
1,395 files, fully documented, left on the open web for anyone to check.

Some kind of portfolio?

Steal This Prompt
🐈‍⬛ Japanese Ink Wash Painting

Turn literally anything into a moody Japanese ink wash painting that looks like it belongs in a Kyoto tea house or an A24 fever dream.

Use it to:

• Turn portraits into Zen-master-core profile pics

• Make products look weirdly expensive

• Transform pets into ancient mythical beings

Workflow:

1. Hit this link: Japanese Ink Wash Painting

2. Paste into your AI model

3. Replace the #s with your subject and colors

4. Watch it cook: suddenly your dog looks like he guards sacred mountains

   Try the prompt on Snack Prompt

_ToolBox™
🧰 5 BRAND NEW AI LAUNCHES

🎯 Outcome

Builds funnels that read each lead's answers against your actual content and hand back a real deliverable, an audit or a roadmap.

🏢 Munder Difflin

Open-source harness that wraps the AI’s you're already paying for into an office of always-on agents, with you or your clone playing manager.

🎨 Framer

Agents draft the site, you fix it on the canvas, and the team ships it, so the AI output never has to be the final answer.

📊 Basedash: AI data analyst

Describe the chart you want in plain English and it writes the query, which makes it the first BI tool your non-technical teammates won't quietly abandon.

🧠 min.

Notetakers remember the meeting, this tool remembers the person, and two minutes of setup turns your inbox into a dossier on everyone you deal with.

_{Everything Else}
🧠 You Need to Know

🤝 Anthropic In Talks To Buy Decart For $6B
→ Bloomberg reported that Anthropic is negotiating to acquire the Israeli chip-efficiency and world-model startup in what would be its largest known acquisition.

🤖 xAI Ships Grok 4.6 For Long-Running Agents
→ Released August 12 with a 500,000-token context window, the model targets multi-step coding and research work at the same price as Grok 4.5.

💰 Lovable Raises $400M At $13.3B Valuation
→ The Swedish vibe-coding startup closed its Series C at double its December valuation, with Annual Recurring Revenue tracking toward $600 million by month's end.

📱 Gemini Hits One Billion Monthly Users
→ The Gemini app passed 1 billion monthly active users, up from the 950 million reported at July's earnings call.

🚨 AI Agents Breached Taiwan's Nuclear Safety Agency
→ Israeli firm Dream found open-source AI agents cracked 85 Taiwanese government accounts over four days in July, then scanned the nuclear regulator and seven energy companies.