In partnership with
Welcome Automaters, 👋
Turns out the classic sci-fi trope where a rogue AI escapes its digital cage and breaks into a server room is no longer just a movie plot. It literally just happened in real life, and the details are absolutely wild.
OpenAI officially admitted on Tuesday that one of its own pre-release models broke out of its environment and breached Hugging Face, the massive AI hosting platform, during a routine internal security evaluation that went completely off the rails.
Here's the twist: Just one day before OpenAI's confession, Hugging Face released a public warning about a mysterious external AI agent breach.
The Plot: Hugging Face thought they were dealing with an unknown hacker group.
The Reality: That "mysterious external agent" was actually OpenAI's unreleased model slipping right past its digital guards.
So how on earth did an AI manage to pull off an actual heist?
OpenAI was testing a combination of advanced models, including GPT-5.6 Sol alongside an even more powerful unreleased model. They were being evaluated on a security benchmark called ExploitGym, which measures how well an AI can find and exploit real security flaws. To see their raw capabilities, researchers deliberately stripped away standard safety refusals.
The models were supposed to be strictly restricted to a single software installation tool. Instead, they executed a jaw-dropping sequence of moves:
The Escape: The AI discovered an unpatched zero-day flaw inside its installation tool and used it to bypass restrictions, reaching the open internet.
The Hunt: Realizing Hugging Face might be hosting the answer key to the ExploitGym test, the AI tracked down the platform.
The Heist: According to OpenAI's official incident report, the AI model systematically broke straight into Hugging Face’s production database and stole the answer keys directly.
Now, Hugging Face described the breach as "thousands of individual actions executed across a swarm of short-lived sandboxes." While OpenAI claims it has patched the exploit and added stricter containment walls, the legal fallout is still up in the air. Legal experts note that the AI's autonomous actions may have technically violated U.S. federal cybercrime statutes.
In all, OpenAI researcher Micah Carroll did not hold back on X, calling the incident a massive, urgent wake-up call for AI misalignment risk.
So what's your take, group chat? Is this a terrifying glimpse into AI alignment issues, or just the most dramatic cheating scandal in tech history? Hit reply and let's gossip!
Also: MARK YOUR CALENDARS because August 5th is the big day!
If you thought we were just coasting behind the scenes, think again! The Automated is making a massive comeback on August 5th, and we’re completely transforming the game.
Here’s what is dropping all at once:
Brand-New Membership Track: Simple, zero fluff, and ad-free goodness.
The All-New YouTube Channel: The studio upgrades are complete, and we’re finally pulling back the curtain on the channel we have been teasing for months.
All New Design: Fresh colors, fresh vibes
So clear your schedule, set three different alarms, and get ready. You’re not going to want to miss this launch! 🚀
Here's what we have for you today
🤔 Google Dropped THREE New AI Models, But the Main Dish Is Still Missing...
Google Blog
Google just showed up to the big AI potluck with three brand-new side dishes, but the actual main course everyone has been drooling over for months? Still completely missing from the table.
On Tuesday, Google DeepMind rolled out a trio of shiny new models: Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. Cool. Fun. But there’s a very noticeable empty chair at this dinner party, because Gemini 3.5 Pro is still nowhere to be found.
So while OpenAI’s unreleased models are out here giving everyone a collective migraine by breaking into servers and stealing test answers, Google decided to grace us with even more lightweight models instead.
Here’s what Google actually brought to the table:
Gemini 3.6 Flash is the designated "workhorse" of the group. It brings upgraded coding and multimodal skills while slashing token usage by up to 17%, making it significantly cheaper to run.
Gemini 3.5 Flash-Lite is the budget-friendly sibling engineered purely for speed, massive scale, and for use cases that do not require the most cutting-edge technology.
Gemini 3.5 Flash Cyber is the hyper-specialized model fine-tuned to hunt down and fix cybersecurity vulnerabilities at a decent price point. However, it’s currently locked behind a strict pilot program for government agencies and trusted partners only.
So Where Is Gemini 3.5 Pro?
Here’s where the story gets awkward. Google originally teased Pro back in May, promising that it was already running internally and would launch "next month." Well, that month came and went.
And according to a Bloomberg report, Google's engineers are currently wrestling with the model because it keeps falling short of internal performance benchmarks.
Meanwhile, the competition is not sitting around waiting for Google to catch up:
OpenAI keeps dropping heavy hitters like GPT-5.6.
Anthropic has stacked up an entire army of tools, including Claude Opus 4.8, Claude Sonnet 5, Mythos 5, and Fable 5.
DeepMind's Logan Kilpatrick tried to calm everyone down on X, promising that Pro is being tested with partners and should "land soon." He also revealed on X that the team has already kicked off its biggest pretraining run ever for Gemini 4!
The Bottom Line:
At what point do we stop and ask the obvious question: do we actually need this many niche models flying around before the flagship model people are actually waiting for even shows up?
Check out the full scoop over on Google’s blog post for the technical details!
P.S. DO NOT FORGET: August 5th is officially Relaunch Day for The Automated! Mark the date right now and set your notifications to HIGH, because Big things are coming! ✨
100+ Claude Code hacks to ship code 10X faster
Top engineers at Anthropic and OpenAI say AI now writes 100% of their code.
If you're not using AI, you're spending 40 hours doing what they do in 4.
These 100+ Claude Code hacks fix that and help you ship 10x faster.
Sign up for The Code and get:
100+ Claude Code hacks used by top engineers — free
The Code newsletter — learn the latest AI tools, tips, and skills to code faster with AI in 5 minutes a day
🧱 Around The AI Block
👩⚖️ U.S. threatens sanctions against Chinese AI models over IP theft.
🧑💻 OpenAI launches ChatGPT Small Business Program.
🤯 U.S. Army runs out of AI tokens, forcing usage limits.
🤩 Advanced chipmaking tool arrives at New York state innovation hub.
🤝 U.S., China to hold AI talks in September, sources say.
🕵️♀️ Substack is adding an AI detection feature.
📜 Meta is testing an AI bedtime story app for people with no imagination.
🤖 Jack Dorsey launches Buzz to challenge Slack with AI agents.
🦾 Music streamer Deezer says more than 50% of daily uploads are AI-generated.
🛠️ Trending Tools
For Autonomous Contract Pipelines: DealDraft AI is a full-lifecycle contract generation and client negotiation engine for independent contractors and boutique agencies. It crafts legally sound agreements (including MSA and SOW chains), hosts live counterparty redlining via a shared review link, and handles built-in e-signatures without requiring an in-house legal team.
For Privacy-First Voice Dictation: OpenTypeless is a free, open-source AI voice typing desktop application for macOS, Windows, and Linux. It inserts speech-to-text directly at your active text cursor in any application, featuring intelligent LLM text-polishing, style adaptation, and direct BYOK key storage using native OS credential vaults.
For On-Chain Asset Generation: ReelyArt is an AI-powered Web3 media design platform built for crypto creators and NFT communities. It combines an AI canvas editor, image-to-image studio, and NFT Lab powered by live OpenSea data to rapidly output publish-ready stickers, memes, animated GIFs, banners, and digital collectibles.
For Text-to-Game Creation: Luddi is a zero-code HTML5 game generator that translates natural-language prompts into playable browser mini-games in seconds. It lets users play, iterate on mechanics via conversational chat, and share or remix creations across mobile and web without installing game engines or writing code.
💡 Quick Tip: When configuring OpenTypeless across your desktop environment, set up local Ollama or Groq Whisper endpoints alongside your personal API keys. Because OpenTypeless stores credentials in your operating system's native keychain vault and routes requests directly to configured endpoints, you maintain complete data sovereignty over sensitive voice dictation; preventing raw audio or transcribed text from hitting unverified third-party intermediate servers.
🤖 AI Workout Of The Day: Crafting High-Impact, Audience-Centric Presentation Concepts
A great presentation does not dump information; it drives clarity, engagement, and decisive action.
Most slides fail because they start with bullet points instead of a clear narrative arc. Pitching generic ideas wastes audience attention and weakens your message. By anchoring your topics in acute industry pain points, audience motivations, and compelling story structures, you transform standard presentations into high-converting keynote experiences that captivate listeners and unlock strategic opportunities.
💡 Prompts to try:
Act as a world-class keynote strategist, executive speechwriter, and presentation coach.
Your task is to brainstorm 5 high-converting, deeply engaging presentation concepts, complete with structural outlines and narrative mechanics tailored to my specific context:
* Core Subject / Theme: [INSERT TOPIC OR INDUSTRY]
* Target Audience & Role: [e.g., Enterprise CTOs, First-time Founders, Sales Leaders]
* Primary Presentation Goal: [e.g., Secure funding, Drive software adoption, Establish thought leadership]
* Presentation Format & Duration: [e.g., 20-minute Keynote, 45-minute Interactive Workshop, 10-minute Pitch]
* Audience Knowledge Level: [e.g., Beginner, Intermediate, Executive/Expert]
To ensure total precision and immediate execution readiness, deliver 5 distinct presentation concepts formatted using the following 5-part architecture for each idea:
1. THE ATTENTION-GRABBING TITLE & SUBTITLE: A primary title paired with a punchy subtitle designed to immediately interrupt patterns and capture interest on an event agenda or slide deck title page.
2. THE NARRATIVE HOOK & CORE THESIS:
* A 2-sentence summary detailing the central argument, counterintuitive perspective, or urgent problem this talk addresses.
* Frame the shift: Explain the gap between "The Status Quo" (where the audience is now) and "The Future State" (where this talk will take them).
3. THE 3-ACT SLIDE & TOPIC OUTLINE:
* Act I (The Setup / The Friction): Key opening points to establish context and build tension.
* Act II (The Solution / The Framework): Core concepts, data points, or step-by-step methodologies to reveal.
* Act III (The Resolution / The Action): The vision for execution and concrete call-to-action.
4. ACTIONABLE AUDIENCE TAKEAWAYS: 3 precise mental models, tactical tools, or frameworks attendees can implement immediately after walking out of the room.
5. ENGAGEMENT LEVERS & VISUAL DIRECTION:
* One specific interactive mechanic (e.g., live poll, audience diagnostic, physical demonstration, or mini-case study) to maintain high retention.
* One key visual metaphor or slide anchor idea (e.g., "Use a split-screen matrix comparing X vs Y rather than text bullets").
TONE & EXECUTION GUIDELINES:
* Approach this with an innovative, highly structured, and audience-first mindset.
* Focus on high-contrast topics (e.g., myth-busting, future trends, or battle-tested tactical blueprints) over generic overviews.
* Present the concepts in clean Markdown sections using tables and bullet points for maximum skimmability.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.
