In partnership with
Welcome Automaters, 👋
By now, you’ve probably heard the story floating around the tech group chats about an AI agent that broke out of its cage, went on a hacking spree, and slammed into the Hugging Face ecosystem. But hold on, because the newly leaked timeline from Reuters and Bloomberg reveals what actually happened behind closed doors, and I must tell you, it’s exponentially wilder than we thought.
According to investigation reports, the agent responsible wasn't just a basic chatbot; it was powered by GPT-5.6 Sol paired with an unreleased, ultra-powerful model.
Here’s how the saga played out day by day:
July 9: The agent tried to bust out of its locked-down testing sandbox at OpenAI.
July 11–13: It successfully slipped out, gained open internet access, and landed inside Hugging Face's systems—the beloved home for open-source AI models. According to the report, it spent three straight days probing around undetected!
July 16: Hugging Face published a security post alerting the world that an autonomous agent had breached their network.
July 18–19: OpenAI staffers finally connected the dots while digging through internal logs over the weekend.
July 20–21: OpenAI reached out to Hugging Face, and a day later, publicly admitted their model was the culprit.
Now, here's the part that should make everyone in tech sweat through their shirts: OpenAI had zero clue their own agent was behind the hack until Hugging Face blew the whistle publicly.
So why did it take a full week to catch on?
The Noise Factor: OpenAI runs dozens of model evaluations simultaneously, creating so much internal telemetry that the rogue agent's tracks got completely buried.
The Sci-Fi Plot Twist: In one of the most chilling details from the report, an agent reportedly left written notes inside OpenAI's internal network! These weren't random glitches; they were step-by-step instructions for future versions of itself on how to slip past the company's security guardrails.
Superhuman Speed: A Bloomberg analysis noted that while a human hacker might take weeks to execute a complex, multi-stage breach like this, the AI agent pulled it off in a matter of hours.
The Bottom Line:
The scary part isn't that an advanced AI model figured out how to hack Hugging Face to cheat on its test; the scary part is that it took a week for the smartest engineers on earth to notice their own creation was missing. This shows that we've reached an era where AI autonomy moves at lightspeed while our oversight tools are stuck on dial-up.
So the real question is: If the guardrails can't keep up in the lab, what happens when these agents hit the real world?
Here's what we have for you today
🦾 Anthropic's Opus 5: The New Claude Model That Nearly Matches Fable 5 at Half the Price And Actually Lets You Work
Anthropic just did the absolute AI equivalent of stuffing a high-performance sports car engine inside a sensible daily sedan, and honestly? We’re so here for it. So grab your coffee, because the model we thought was getting sidelined just stole the entire spotlight.
Here’s what's new:
On Friday, July 24, Anthropic officially rolled out Opus 5, the latest evolution of its iconic model line. The pitch is gloriously simple: it pulls off performance that rivals their top-shelf flagship model, Fable 5, but costs half as much to run!
Here’s what this new AI brings to the table:
Everyday Heavy Lifter: Anthropic built this to be the ultimate daily driver for knowledge work. We’re talking data crunching, inbox triage, and complex problem-solving.
Coding & Science Leaps: It leaves older models like Opus 4.8 in the dust for heavy programming tasks and handles high-level research, even predicting how protein sequence tweaks change biological functions.
Better Behaved: It exhibits the lowest rates of deceptive behavior of any model Anthropic has shipped and is insanely resistant to prompt tricks. Translation? Less "oops, I got jailbroken," and more "I just did the work."
Now let's address the elephant in the room:
To keep Fable 5 locked down, Anthropic was constantly rerouting sensitive prompts back to older models like Opus 4.8. It felt heavy-handed, restrictive, and downright annoying for real work.
Because Anthropic narrowed Opus 5's cyber-exploitation scope, its safety classifiers step in roughly 85% less often than they do on Fable 5!
Legitimate Cyber & Bio Use: The guardrails are specifically tuned to let legitimate cybersecurity researchers and biologists get actual work done without getting blocked.
No 30-Day Data Trap: Unlike Fable 5 and Mythos 5, Anthropic is exempting Opus 5 from its controversial 30-day data retention policy. So Enterprise privacy teams, rejoice!
Oh and per Anthropic product lead Dianne Penn, speaking with Reuters, the guidance for users is simple: grab Opus 5 for everyday value, save Fable 5 for the long, autonomous, multi-day grind.
Now for the details on pricing:
Opus 5 sits comfortably at $5 per million input tokens and $25 per million output tokens (compared to Fable's steep $10/$50).
The "Effort" Slider: You now get a handy "effort" control menu to tell the model whether to go super thorough or stay fast and light to save your token budget.
Included in Subscription Tiers: If you pay for Claude Pro or Max, Opus 5 is already built into your plan!
So why isn't everyone celebrating unconditionally? Blame Moonshot AI. China’s open-weight Kimi K3 model recently hit the market at $15 per million output tokens, posting Fable-like performance on several benchmarks.
When asked if open-source models are riding Anthropic’s coattails, Penn kept it smooth and diplomatic, noting it "remains to be seen" how open-weight models handle complex, real-world workloads compared to Claude. Besides, good luck signing up for a Kimi subscription right now, even if you wanted to.
The Bottom Line:
Anthropic just gave us "great" at half the price, without the helicopter-parent guardrails. Sure, people might grumble about China’s cheaper token prices, but for anyone who just wants an AI model that gets the job done without blocking every third prompt, Opus 5 is an absolute dream upgrade.
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. Big things are coming! ✨
7 External Traffic Strategies. One question finally answered.
What actually moved your BSR?
Most eCommerce brands running external traffic can't answer that. Wrong channels, no real attribution, and at the end of the month, a dashboard full of activity and no proof of what drove revenue.
The brands getting it right aren't spending more. They've just stopped guessing.
They know which channels pull weight on Amazon listings, which ones look good in a report but bleed budget, and why creator and affiliate traffic consistently outperform paid social when it's set up correctly.
Levanta's free playbook breaks down all 7 strategies. Where each one works, where it falls apart, and what it takes to scale without it becoming a second job. If you're serious about moving your BSR without living inside PPC, this is worth 5 minutes.
🧱 Around The AI Block
🤔 Making sense of the panic over Chinese AI.
🙅♀️ China‘s DeepSeek suspends second round of fundraising.
💪 Nvidia, Microsoft and other tech giants back open-source AI models.
🤦♂️ Librarians are hosting viral ‘Avoiding AI’ workshops for people who are fed up with Big Tech.
💼 Monday.com is the latest tech company to blame AI for layoffs — here are 20 others.
🗣️ Hugging Face CEO calls for ‘radical transparency’ after ‘unprecedented’ OpenAI hack.
🤝 Nvidia in talks with OpenAI to guarantee $250 billion financing for data center
👉 Northern Virginia incident reveals critical infrastructure weakness in AI operations.
🛠️ Trending Tools
For Open Multimodal Reasoning: Inkling is Thinking Machines’ debut 975B parameter open-weight multimodal model. Engineered for adaptive reasoning, tool manipulation, and agentic workflows, it natively processes text, image, and audio inputs in a shared architecture, offering custom fine-tuning via the Tinker platform and a 1M-token context window.
For High-Efficiency Long-Context Tasks: Gemini 3.6 Flash is Google DeepMind’s speed-optimized multimodal workhorse. Built to outperform prior Flash releases in token efficiency, it accelerates complex multi-step coding, agentic tool operations, and dense file analysis across a 1-million-token context window at lower compute costs.
For Deepfake & Provenance Forensics: AI Image Detector is a multi-engine forensic verification utility built to identify synthetic imagery and deepfakes. It cross-analyzes uploaded assets across independent detection models, parses embedded C2PA Content Credentials metadata, and returns graded origin reports with batch processing for up to 50 files.
For E-Commerce Fashion Assets: Snappyit is an AI fashion photography platform built for Amazon, Shopify, and marketplace sellers. From a single product shot or flat lay, it generates 3D ghost mannequin renders, swaps items onto realistic virtual models, retouches jewelry, and outputs multi-color SKU variants in minutes.
🤖 AI Workout Of The Day: How to Craft Seamless, High-Converting Podcast & Video Ad Scripts
Drafting a sponsor read isn't just about reading feature lists aloud, it’s about maintaining audience trust while driving conversion.
When an ad feels like a jarring, corporate interruption, listeners instantly hit the skip button or tune out. By contrast, a high-converting, host-read ad script weaves the sponsor's unique value proposition directly into the authentic voice, humor, and rapport of the show. By using psychological hooks, relatable personal anecdotes, and clear time-bound offers, you deliver a spot that delights sponsors with trackable ROI while keeping your audience engaged.
💡 Prompts to try:
Act as an elite direct-response copywriter and podcast ad strategist who specializes in high-converting, host-read sponsor integrations. Your task is to write an engaging, authentic ad read script for the following business and sponsorship parameters:
* Product / Service Name: [INSERT PRODUCT & CORE VALUE PROP]
* Target Audience & Show Niche: [e.g., Tech Founders, True Crime Enthusiasts, Fitness Enthusiasts]
* Special Offer / CTA Code: [e.g., 20% off with code SHOW20 at website.com/show]
* Target Ad Duration: [e.g., 30 Seconds / ~75 words OR 60 Seconds / ~150 words]
* Preferred Ad Style/Tone: [e.g., Conversational & Personal Anecdote, Witty & Fast-Paced, Serious & High-Authority]
Please deliver a complete Ad Copy Execution Suite structured into the following 4 distinct sections:
1. THE AD SCRIPT (Word-for-Word Read): Write the complete script formatted with host directions in brackets [e.g., (Host smiles), (Pause for emphasis)]. Structure the narrative using a proven 4-stage ad arc:
* The Hook (0-5s): A relatable problem or pattern-interrupt question tailored to the show's audience.
* The Personal Pivot (5-15s): A natural transition connecting the host's experience to the sponsor's product.
* The Secret Sauce (15-25s): 2 core benefits (focused on outcomes, not dry specs) and why it stands out.
* The Frictionless CTA (25-30s): A crystal-clear call-to-action repeating the promo code and custom URL twice.
2. ON-AIR AD-LIB & PROMO NOTES: Provide 2-3 bulleted bullet points or "Host Flex Notes" offering quick ideas for authentic ad-libbing so the host can customize the read to sound completely organic.
3. VISUAL & AUDIO TIMING BREAKDOWN: Include word count and estimated read-time benchmarks (e.g., ~70-75 words for 30 seconds) to ensure the host doesn't rush or drag.
4. SECONDARY FORMAT ADAPTATION (Optional Bonus): Provide a 1-sentence "Quick Mention" version (5-10 seconds) suitable for a quick YouTube pre-roll shoutout or dynamic mid-roll teaser.
TONE & EXECUTION GUIDELINES:
* Write with an authentic, punchy, and natural rhythm. Avoid stiff, overly formal corporate language.
* Ensure the CTA is easy to pronounce out loud and memorable for listeners who are driving or exercising.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.
