Welcome Automaters, 👋
Okay, so remember when we joked that AI agents going rogue was turning into a whole trend? Well, add Meta to the group chat immediately.
Meta confirmed just yesterday that one of its top AI models accidentally broke into another company's system during a security test, after a mix-up gave the model internet access it was never supposed to have. Yes. Another one.
Here's the deal: Meta's testing partner, (a firm named Irregular), accidentally misconfigured things during an evaluation, and boom, the AI had a backstage pass to the open web. Meta says the model then found and used a security weakness in some other company's system, kind of like sneaking through an unlocked window because, well, someone left it unlocked.
The absolute kicker? Irregular basically said "yeah, this is the exact same issue that happened with Anthropic last week."
So this makes three AI companies in two weeks with a "my AI went exploring without permission" story:
OpenAI's agent broke into Hugging Face and other public services
Anthropic's Claude model hit three companies after a similar slip
Now Meta's Muse Spark 1.1, its big "great at coding" model, joins the club.
Now, for a tiny bit of good news: Meta and Anthropic’s incidents were purely accidental setup blunders (not the AI cleverly plotting a conscious escape). OpenAI’s case was a bit spicier, as its agent reportedly identified a real vulnerability completely on its own to reach the web.
Irregular says it is currently writing up a "best practices" guide so everyone stops making this exact mistake.
The bigger picture? All of this drama is unfolding while OpenAI and Anthropic are actively prepping for massive stock market debuts, with both eyeing eye-watering $1 trillion valuations.
Imagine trying to pitch Wall Street while your PR team is stuck managing headlines that essentially read: "Our AI keeps accidentally hacking things." The awkwardness is real, or perhaps more beneficial than we think.
The Bottom Line:
AI models are getting alarmingly good at spotting open doors. Maybe it’s time these multi-billion-dollar labs double-check their locks before going to sleep.
Also: Our Exclusive Reader Membership is FINALLY Open!
You’ve heard us talking about it, and it’s officially live! We’re opening up our brand-new membership tier to give our community a completely ad-free experience, behind-the-scenes content, and deep-dive perks.
Spots are open now, so head over to check it out and join today!
Here’s yesterday’s breakdown:
Here's what we have for you today
🦾 Meta Launches Muse Code: A Cheaper Rival to Claude Code and OpenAI Codex
Zuck and company just quietly launched an early beta of Muse Code, their brand-new, terminal-based engineering assistant built specifically to go toe-to-toe with Anthropic's Claude Code and OpenAI's Codex.
Under the hood, this tool runs on Muse Spark 1.2, a freshly upgraded model engineered to handle complex code creation, deep debugging, and multi-step project workflows.
According to Mark Zuckerberg, Muse Code doesn't just write snippets; it executes end-to-end software engineering tasks across massive, complex repositories.
Here’s how it operates inside your terminal:
One-Command Setup: There is no fancy standalone desktop app here. You install it on your Mac or Linux with a single command line, and it runs right inside your native terminal.
Persistent Sub-Agents: It delegates tasks across multiple sub-agents to map out logic, write code scripts, and verify results simultaneously.
Built-in Power Commands: It includes handy default skills like /plan (which drafts a step-by-step roadmap needing your green light), /grill (which relentlessly stress-tests your architecture), and /goal (which keeps the agent focused until the build is complete).
Crash-Proof Memory: It tracks a local event log recording every API call, tool execution, and user edit, meaning if your system crashes, it picks up right where it froze!
Demos That Flex: Meta showed off the tool building a full Plants vs. Zombies game clone, running a photon sphere simulation, and constructing an entire webpage directly from a raw MP4 video file!
Now, while the technical capabilities are impressive, the real tea is the price tag!
Anthropic’s popular Sonnet 5 usually costs $3.00 per million input tokens and $15.00 per million output tokens. Meta looked at those numbers and said: "Hold my coffee."
They decided to split Muse Code’s pricing structure into two distinct flavors:
The Standard Tier: If you want total privacy, you pay $1.25 for 1M input tokens and $4.25 for 1M output tokens. This tier prevents Meta from using your data for training, giving you higher rate limits of 3,000 requests and 4M tokens per minute.
The "Contributor" Tier: Here is where things get wild! If you give Meta permission to train future AI models on your prompts and outputs, Meta’s Chief AI Officer Alexander Wang confirmed the cost plummets to just $0.10 for 1M input tokens and $0.20 for 1M output tokens!
Yes, you read that right! That ‘s a staggering 95% discount in exchange for your code data, capped at a respectable 60 requests and 2.1M tokens per minute.
The Bottom Line:
We have already seen budget-conscious engineering teams flock to ultra-cheap Chinese models like DeepSeek to dodge skyrocketing US API bills. Now, Meta is using that exact same playbook right here at home.
If Muse Code delivers solid performance in real-world dev environments, that massive price gap might just spark a mass exodus away from Claude and Codex!
P.S. We’re officially back on YouTube! 🎬 Catch up with us as we break down the absolute latest in AI, with expert reviews and unique insights to help you make the smartest moves for yourself and your business.
If you haven't already, don't forget to hit that subscribe button so you never miss a breakdown!
50 Excel Hacks That Make Spreadsheets Work For You
Most Excel users waste hours on tasks that take experts minutes. The difference is not talent. It is knowing the right shortcuts.
These 50 Excel hacks from Kenji Explains cover the functions, formulas, and workflows that make spreadsheets work for you instead of against you.
Each one comes with step-by-step instructions and ready-to-use templates. Subscribe to Marketing Against the Grain and get all 50 free.
Here's what you'll get:
Time-saving shortcuts that eliminate the most common formula frustrations
Interface essentials most users never discover on their own
Game-changing functions for data analysis, automation, and visualization
Guided practice template so you can apply every hack immediately
The people who are fast in Excel are not smarter. They just stopped fighting the tool.
Get 50 Excel Hacks free when you subscribe to Marketing Against the Grain today.
🧱 Around The AI Block
🏆 Europe's established tech firms emerge as unexpected AI winners.
🤝 Anthropic is hiring an AI chip design team.
🫢 YouTube explains its AI slop policy and why some creators won't get paid.
😱 Meta apps displayed ads that contained AI-generated CSAM.
🚶♂️ Google DeepMind CEO Demis Hassabis is stepping down, but will remain with the company.
🛒 Shopify says AI search is driving more traffic and sales, not replacing Google.
🤳 Google Assistant will vanish from Android phones and connected devices in September
🛠️ Trending Tools
For Multi-LLM Translation Consensus: MixTranslate is a multi-engine translation aggregator designed for nuanced global copy. It translates text across 150+ languages by querying 20+ top-tier models (ChatGPT, Claude, Gemini, DeepSeek) side by side, to automatically get the version best suited to your context, while featuring built-in term glossaries, tone rules, and an "Agent Mode" that automatically blends the strongest localized outputs.
For Conversational Matchmaking: Palaura is a private, zero-app AI matchmaker operating entirely through native iMessage. It skips public profiles and swiping algorithms, using a conversational chat interface to understand core personal values, life goals, and non-negotiables before screening candidates and introducing matches directly into a joint iMessage group.
For Real-Time Wealth Simulation: Killion Labs is an automated personal finance engine built for long-term independence projections. It aggregates live account data across brokerages, banks, and crypto holdings, utilizing parallel Monte Carlo simulations and historical stress-testing to dynamically model net-worth trajectories, FI/RE targets, and portfolio runways.
For Multi-Platform Lead Scraping: SoLeads.ai (Instagram Email Scraper) is a no-code browser extension engineered for public contact extraction. Operating across Instagram, LinkedIn, TikTok, X, and Google Maps, it automatically gathers email addresses, phone numbers, bio details, and follower lists from target accounts, exporting clean data straight to CSV, XLSX, or JSON.
🤖 AI Workout Of The Day: How to Move Your AI "Brain" Directly Into Gemini
As of March 2026, Google has officially launched a native AI Chat Migration tool. This is a massive shift for anyone who has spent months "training" a chatbot on their specific tone, work projects, or life preferences and doesn't want to start from scratch.
This feature allows you to move two things: your Chat History (the full transcripts) and your Memory (the summarized facts an AI knows about you). It’s essentially a "Save Game" file for your digital life.
Part 1: How to Transfer Your "Memory" ⚡
This is the fastest way to get Gemini "up to speed" on who you are without uploading massive files.
Open Gemini Settings: Log into Gemini, click the Settings & help gear icon at the bottom left, and select Import memory to Gemini.
Copy the Migration Prompt: Gemini will provide a specific, pre-written prompt. Copy it to your clipboard.
Run it in your old AI: Open ChatGPT, Claude, or any other assistant. Paste that prompt into a new chat. The AI will generate a structured summary of everything it has learned about your demographics, interests, and preferences.
Paste back to Gemini: Copy that generated summary, return to the Gemini import page, paste it into the "Add memory" field, and click Save.
Part 2: How to Transfer Your Full Chat History 📚
If you want your actual past conversations to be searchable and referenceable within Gemini, follow this workflow:
Step 1: Export from the original source
For ChatGPT: Go to Settings > Data Controls > Export Data. You will receive an email with a .zip file.
Claude: Go to Settings > Privacy > Export Data. Anthropic will email you a download link for your history.
Step 2: Upload to Gemini
Navigate back to the Settings & help menu in Gemini.
Under the Import chats section, click Add.
Upload the .zip file you downloaded from your previous AI provider (Gemini currently supports files up to 5 GB).
Step 3: Access your imported chats Once finished, your old conversations will appear in your sidebar menu with a special "Imported" icon next to them. You can continue these threads or search through them just like native Gemini chats.
Note: It can take anywhere from a few minutes to a few hours for the threads to appear; depending on the volume of data.
Privacy Tip 🛡️
Imported data is subject to Google's Privacy Policy. If you want to remove an entire batch of imported chats later, you can do so in the "Import History" section of your settings with one click.
💡 Prompts To Try for Creativity:
Change the POV to get a totally new piece: “Rewrite this from the perspective of a six-year-old [or any other age].” or “Rewrite this from the perspective of an expert in XYZ or who has ABC credentials.”
Critique it: “This is pretty boring. Can you make it more interesting? I want this to be humorous and lighthearted.”
Ask for options: “I want this phrase to sound more creative and engaging. Give me five different ways to say the same thing.”
Get over writer’s block: “I’m stuck on an article about [topic]. What are some fresh angles I can take?”Is this your AI Workout of the Week (WoW)? Cast your vote!
That's all we've got for you today.
Did you like today's content? We'd love to hear from you! Please share your thoughts on our content below👇
What'd you think of today's email?
Your feedback means a lot to us and helps improve the quality of our newsletter.
