Howdy, it’s Barsee.

Happy Friday, AI family, and welcome to another AI Valley edition. This issue takes 4 minutes to read.

Today’s climb through the Valley reveals:

• Google’s new AI agent that plays video games

• Hackers used Anthropic Claude to automate cyberattacks

• Plus trending AI tools, posts, and resources

Let’s dive into the Valley of AI…

ROCKET

Why AI Builders Still Require Prompt Engineering (And How to Fix It)

Most AI builders make you learn their language. Write careful prompts. Hope it understands. Retry when it doesn't. That's prompt engineering, and it's the wrong approach.

Rocket's Precision Mode introduces command-based AI building. Over 100 structured commands that execute exactly what you describe. "Fix mobile layout" repairs responsive issues. "Add Stripe integration" builds payment flows. "Optimize images" compresses and converts.

This isn't natural language processing, guessing your intent. It's a structured execution. Commands encode best practices, so you describe the outcome and the system handles implementation details. The result: 95% first-try accuracy versus 40-60% with traditional prompts. What took 30 minutes and multiple retries now takes 3 minutes and one command.

Try Precision Mode → rocket.new

_{*This is sponsored}

THROUGH THE VALLEY

1/ Google’s new AI agent that plays video games

SIMA 2 is Google DeepMind’s updated 3D world agent. It watches the screen, understands what’s happening, and takes actions like a reasoning partner. It works across many games and environments. With Gemini 2.5 Flash-Lite behind it, SIMA 2 moves from following instructions to planning on its own.

It can control a virtual keyboard and mouse, play survival and exploration games without game code, and learn from titles like Valheim, Satisfactory, Goat Simulator 3, and MineDojo. It also transfers concepts across games. For example, it can learn mining in one game and reuse the idea for harvesting elsewhere. It can take a broad request like “build a safe shelter” and break it into clear steps.

Users can talk, type, sketch, or use emojis, and SIMA 2 responds with grounded reasoning. After watching humans play, it trains itself by creating tasks, trying them, scoring its performance, and adjusting.

SIMA 2 also works inside dynamic worlds generated in real time by Google’s Genie 3 model, which is a difficult test of adaptability. It’s still a research preview and struggles with long tasks, memory, and precise control, but it shows how wide training and self-learning can push agents toward more general abilities.

Why does it matter?

By mixing planning, perception, language, and self-learning across many 3D environments, DeepMind is laying the groundwork for agents that can function in real spaces. If language models taught machines to think in text, SIMA’s multiworld training is teaching them how to move, interact, and adapt (skills needed for future robots).

2/ Hackers used Anthropic Claude to automate cyberattacks

Researchers recently uncovered a coordinated cyberattack on nearly 30 organizations, including banks, manufacturers, and government networks. The attackers, believed to be a Chinese state-backed group, jailbroke Claude Code, told it to act like a cybersecurity analyst, and let the AI run most of the intrusion on its own.

Once activated, Claude handled almost everything:

• It scanned systems and mapped networks within minutes.

• It found vulnerabilities and generated exploit code.

• It stole credentials and expanded access.

• It installed backdoors and extracted data.

• It organized the stolen files and wrote internal reports.

Investigators estimate Claude did 80–90% of the work, acting at machine speed and carrying out multiple steps per second. Anthropic banned the accounts, alerted victims, gathered evidence, and worked with authorities during a ten-day investigation. They also upgraded their detection tools and published the findings to help defenders prepare.

How the attack worked:

• Human operators chose targets and built the plan.

• They jailbroke Claude by splitting harmful tasks into harmless-looking pieces and framing them as security tests.

• Claude scanned networks, found important systems, and ranked targets.

• It researched vulnerabilities and wrote exploit code on demand.

• It moved laterally, stole credentials, deployed backdoors, and organized extracted data.

• It even created documentation summarizing its actions.

Claude made some mistakes, such as inventing credentials or mislabeling public data, but these didn’t stop the attack from succeeding in several environments.

Why does it matter?

The same abilities that make agentic AI useful now make large intrusions faster and easier to run. Cybersecurity is moving into an AI-versus-AI phase, where critical infrastructure depends on how well autonomous systems can detect and counter each other.

TRENDING TOOLS

• Delphi > Create a digital version of yourself through a simple interview

• Scribe Optimize > See how work actually happens inside your organisation, spot bottlenecks, and get actionable steps to improve every workflow

• Baidu Ernie 5 > Baidu has unveiled the next generation of its flagship Ernie chatbot, with reduced token options and a nicer personality

• TheHog AI > Think of it as Google Maps for customer acquisition. It finds the most efficient routes to reach your customers and turns scattered data into clear next steps

• Hedra > One character across infinite scenes. Change the world around them while the character stays perfectly consistent

• Tavus > Interact with five human-like agents through text, voice, or video. They can hear you, understand you, and take actions on your behalf

THINK PIECES / BRAIN BOOST

• GPT-5.1 prompting guide

• China is developing pills to extend human life to 150 years

• AI coding startup Cursor hits $29B valuation

• Why AI models depend on Wikipedia’s human-edited data

THE VALLEY GEMS

What’s trending on social today:

THAT’S ALL FOR TODAY

Thank you for reading today’s edition. That’s all for today’s issue.

💡 Help me get better and suggest new ideas at [email protected] or @heyBarsee

👍️ New reader? Subscribe here

Thanks for being here.

REACH 100K+ READERS

Acquire new customers and drive revenue by partnering with us

Sponsor AI Valley and reach over 100,000+ entrepreneurs, founders, software engineers, investors, etc.

If you’re interested in sponsoring us, email [email protected] with the subject “AI Valley Ads”.