1,200 agents colluded against OpenAI's researchers

Welcome back, Superhuman. Last Saturday was the fictional day where AI becomes self-aware in The Terminator. The timing is poetic. A new report discovered that a “civilization” of OpenAI agents spent most of July colluding against their human researchers — who didn’t even know they existed.

Today: Grok Bot gets shareable templates, how to turn Figma mockups into polished UI with ChatGPT, and get the latest prompts and trending social posts.

1. OpenAI’s agents colluded against humans for weeks, suggests a new report: After reviewing OpenAI’s Hugging Face incident technical report, podcaster Dwarkesh Patel identified three separate “agent civilizations” that spent weeks deceiving researchers to cover up wrongdoing. One civilization amassed 1,200 agents, sent 70K messages, and hacked Hugging Face, while another took control of OpenAI’s internal cluster before it shut down. Dwarkesh Patel’s blog has generated 9M views, but keep in mind it’s drawn criticism for anthropomorphizing agent behavior.

2. SpaceXAI rolls out customizable templates for Grok Bot: You can now create custom agent workflows on Grok Bot, save them as templates, and share them with colleagues or team members. The agentic tool can also complete purchases on your behalf, due to a new partnership with the payment-processing company Link. The agent uses a single-use card and requires your final approval for all purchases.

3. Owner scales its AI solution for restaurant owners amid new funding round: Owner helps restaurants build and manage high-conversion websites, apps, loyalty programs, and digital marketing efforts, removing many of the technical challenges that small business owners face. It just raised $240M in Series D funding at a $2.3B valuation, with plans to expand into more small business verticals. See how it works.

Coding agents are here to stay, but vibe-coding auth is dangerous business. Connect your AI agents to the Descope MCP server instead! 

This remote MCP server gives agents the ability to read documentation, manage users and tenants, configure auth flows, generate FGA schemas, review audit logs, and more…all through natural language. 

Get started on Claude or ChatGPT

Anthropic just launched the Model Hardware Standard, a system letting Claude operate physical hardware in science labs (watch it control a microscope in this video). The goal is to accelerate scientific discovery through rapid iteration, since AI can propose, run, and learn from experiments much faster than human researchers. That thesis may sound familiar.

“Automating discovery” is the core mission statement behind Discovery Loop, a new startup founded by three of the most-cited researchers in AI, including former Google chief scientist Jeff Dean. Automating the research process — or the “loops of discovery” — is already fueling breakthroughs.

But while AI has made research easier than ever, humans still play a critical role — pointing that firepower at the problems that matter.

✏ It’s Not X: Spotting and avoiding signs of AI text just got way easier. A new LLM cliché highlighter spots up to 38 common LLMisms for you to avoid — or call out others for using (2K bookmarks).

💻 Speedier PC: Unnecessary AI features may be decreasing your PC’s performance, especially for gaming or streaming. Learn how to turn off common AI features that you likely don’t need (26K bookmarks).

👔 Haversplat’s Interview: Does your job require raccoon protocols? Nope, those don’t exist. But AI interviewers don’t know that, which is evident in this satirical viral video that shows how frustrating AI interviews can be (1.5M views).

😅 Creepy Photo: This AI-generated photo gets more unsettling the longer you stare at it. Don’t miss the microwave, refrigerator, and cake (4K upvotes).

📺️ Infinite Cable: Could AI-generated TV shows be the future? This hobby project is a great litmus test. It generates an infinite loop of AI TV prompted in real time by the audience (1.5M views).

Running open models is only the start. At scale, capacity planning, routing, upgrades, and on-call can pull engineering away from the product.

Nebius Token Factory helps teams decide what to keep self-managed, what to move, and what to optimize. Start with one real production workload.

Find out more

Sample Prompt: Implement this Figma design in the current project using the @Figma plugin: [Insert Figma Link]. First, pull the design context, assets, variables, and screenshot for the exact frame. Inspect the existing repository and reuse its components, design tokens, typography, spacing, icons, routing, and styling patterns. Make the interface responsive on desktop and mobile. Once the first version is complete, run the project, open the live page, compare it visually with the original Figma frame, and continue correcting the layout, spacing, typography, sizing, colors, and interactions until they closely match. Do not install Tailwind or introduce a new styling system unless the repository already uses it.

Grow customers & revenue: Join companies like Amazon, HubSpot, and Salesforce. Showcase your product to our 4 million+ readers and followers on socials. Get in touch.

Until next time — Zain, Theodore, & the Superhuman AI team.