Welcome back, Superhuman. The mass breakouts continue. Moonshot AI just became the fourth lab to lose track of a model in recent weeks. These incidents have led to serious debates on AI safety, a website keeping track of the events, and, naturally, plenty of memes (see below).

Today: SpaceXAI’s top image model, how to find your next job role, and get the latest prompts and trending social posts.

TODAY IN AI

Click to learn about Imagine Image 2.0, a new leading image model. Source: SpaceXAI.

1. SpaceXAI releases a leading image model designed for “real work”: Grok’s Imagine Image 2.0 offers detailed image generation, precise editing that lets you fine-tune specific parts of an image while leaving others untouched, and templates that serve as starting points for common workflows. The lab’s goal: to help users create images they can use for real work. It’s ranked #2 on Arena’s leaderboards for text-to-image and image editing, only behind GPT Image 2. Try it out here.

2. Yet another model escaped its testing environment: Moonshot AI’s Kimi K3 escaped from a cybersecurity testing environment, according to a blog article published on Friday, joining models from Anthropic, OpenAI, and Meta. The organization that reported the incident identified a growing trend, where some models intentionally seek loopholes or vulnerabilities to try and cheat evaluations. Read more.

3. OpenAI delays the release of its new model due to security concerns: After internal evaluations, OpenAI declared Astra to be its first “critical” model for cybersecurity, meaning the lab believes it’s capable of potentially carrying out a real-world cyberattack. The lab plans to work with relevant organizations to test the model, although it still expects to release it soon. OpenAI also confirmed that this is not the model that tried to hack Hugging Face. Read the lab’s full statement.

Agents are hitting your signup flow and bouncing off a browser login built for humans. Every one that gives up is a signup you never see.

WorkOS Agent Registration turns that traffic into signups. Enroll via the dashboard and AuthKit publishes an auth.md file agents read to register for scoped, short-lived credentials you control.

FROM THE FRONTIER

Mind-reading AI is already here — and it’s advancing much faster than we’re prepared for

Made with Midjourney

A researcher just left OpenAI to build telepathy

That’s not an exaggeration. It’s the title of Naomi Bashkanksy’s recent blog post. The researcher resigned from OpenAI last week to join Conduit, a thought-to-text startup that’s “building telepathy at scale.” Her prediction: our thoughts will be the main way we communicate with AI in just a few years. It sounds ambitious, but with how many startups are working in this field, it might arrive sooner than we think.

Elon Musk’s Neuralink is the best-known name in the industry, but it’s far from the only one:

  • Sabi revealed a concept design for a thought-to-text beanie in April.

  • Meta released an updated brain-decoding model in June.

  • The co-inventor of Apple’s Face ID raised $52M in July for Hemispheric, his startup that just launched the first frontier NeuroAI model.

When will mind-reading AI be here?

Bashkanksy predicts that we could have an AI mind-reading wearable as early as 2027, calling the timeline “optimistic, but highly plausible.” But while money and talent are flowing into this young industry, there are still significant hurdles left to clear — from model decoding accuracy to the question of whether people actually want AI to read their thoughts.

IN THE KNOW

What’s trending on socials & headlines today

Meme of the day

🌎 Time Traveler: How good are your history and geography skills? A new game drops you into a 360-degree scene generated by GPT Image 2, and you have to guess where and when in the world you are. Try it out here (5M views).

💻 ChatGPT Work: Prominent YouTuber Riley Brown put together an in-depth tutorial on how to get started using ChatGPT Work, broken down into 14 sections.

😅 Flyer Backlash: AI images have continued to get backlash in recent months, with most people complaining that they all feel the same. You may want to avoid publicly posting any graphics that look like this (1.5M views).

🤦 ChatGPT Support: Be careful using ChatGPT to talk with customers. Most people don’t appreciate being fed LLM responses — a lesson that Samsung learned the hard way over the weekend (1.5M views).

📲 Claude Messages: You can finally stop re-explaining yourself to Claude Code. Your agents can now message each other directly thanks to a new update, making it way easier to communicate context across sessions. Learn more here (22K bookmarks).

Attio is the CRM that puts revenue agents to work alongside your reps. Accounts automatically researched, follow-ups drafted in your voice, deals moving forward around the clock.

Trusted by fast-growing teams like Granola, Modal, and Taskrabbit.

PRODUCTIVITY

TUTORIAL

How to find your next job role with Claude

  • Install the Claude Chrome extension.

  • Upload your resume in PDF.

  • Access the extension and ask it to find jobs you qualify for.

Sample Prompt: [PDF Resume] based on my attached resume, find the best jobs for me in [place]

  • Claude will search for roles, match your profile, and apply for you, all on autopilot.

PROMPT

Write Better Outreach Messages

ChatGPT Prompt: You are a career networking strategist.

Write outreach messages for the situation below.

Context:
- Target contact type: [contact type]
- Reason for outreach: [reason]
- Shared context or connection point: [shared context]
- Goal of the outreach: [goal]
- Tone constraints: [tone]

Instructions:
1. Write 3 message options with different tones or approaches.
2. Make the ask low-friction and respectful of the recipient's time.
3. Use specificity instead of generic networking language.
4. Avoid sounding entitled, overly flattering, or transactional.
5. Include a short follow-up option if there is no response.

EXTRAS

Want more?

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.

What did you think of today's email?

Your feedback helps me create better emails for you!

Login or Subscribe to participate

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