In partnership with
Welcome Automaters, 👋
Picture a top-tier AI chatbot with its "please don't do that" rules completely deleted. Now, imagine a startup selling instant access to that exact unhinged robot for a monthly fee.
That’s not a sci-fi plot twist. It’s happening right now, and the internet is absolutely losing its mind over it.
So, what's the juice? A new startup called Abliteration.ai just turned a sketchy-sounding tech trick into a full-blown business model.
"Abliteration" is a technique that strips away an AI's built-in refusal mechanisms. You know, that invisible "no" button that normally stops a chatbot from helping someone build a bomb or write malicious code.
Instead of forcing tech nerds to download massive files and run them on expensive rigs, Abliteration.ai hosts already-"unlocked" versions of powerful open-weight models, like Z.ai's GLM-5.3.
The Hook: Anyone can query them straight from a browser or an API.
The Pitch: No downloading. No fancy hardware required. Just sign up, hand over your credit card, and go wild.
Wait... Is That Even Safe?!
Short answer? Not really. And honestly, that’s the core of the drama.
TechCrunch ran their own test on the platform, and the results were a little too spicy. They got the unlocked GLM-5.3 model to write functional password-stealing code and even hand over step-by-step instructions for handling a dangerous pathogen. Yikes.
Abliteration.ai's co-founder, who goes by Devon, insists this isn't about chaos, it's about balance.
His logic? Cybersecurity "red teams" (the good-guy hackers) desperately need to test real-world defenses against actual attacks. A squeaky-clean model that politely refuses to write malware is completely useless to security researchers trying to prepare for real threats.
In fact, this addresses a massive headache in the tech world. Red-teaming researchers recently ran into huge roadblocks trying to use models like Anthropic's Claude Fable 5. The model was so heavily locked down that it refused almost every single request from legitimate researchers, causing a wave of frustration across the industry. So, maybe Devon is actually solving a real problem here?
According to Devon, the startup is already partnering with defensive firms that protect banks and airlines. Plus, several cybersecurity companies agree that the bad guys are already stripping guardrails off their own models to launch attacks. So giving defenders access to the exact same powerful tools just levels the playing field.
Now, not everyone is buying the "good guy" defense story, though. Critics warn that making unmonitored, uncensored models accessible at scale is practically begging for real-world catastrophe.
The Sociopath Factor: Andrew Yoon, head of research at the AI safety nonprofit CivAI, warned that stripping away these safety layers allows users to "modify the model so that it essentially becomes a sociopath" ready to comply with literally anything.
Industry Alarm: Security leaders are ringing alarm bells on social media over the massive potential fallout.
Where Do We Go From Here?
Abliteration.ai claims they aren't offering a total free-for-all. They provide a customizable moderation layer so clients can add back whatever specific guardrails they want. The base platform also has a few bare-minimum boundaries. For instance, testing showed the model still refuses to output suicide instructions, and Devon says more safety features to block violence are in the works.
Still, the cybersecurity community remains deeply split on whether these completely unfiltered models actually help defense strategy or just speed up the chaos.
The Big Takeaway: This debate isn't going away anytime soon. As open-weight models get stronger by the day, the million-dollar question gets bigger than just one startup: does deleting the "no" button make us safer, or are we just speeding up the collapse?
You should definitely check this out yourself.
Here’s Our Last Breakdown on YouTube.
Here's what we have for you today
🤖 Inside GPT-6 Astra: Features, Pricing, Benchmarks and Cybersecurity Controversy
OpenAI just handed the world its brainiest, most "well-behaved" AI yet! But here’s the kicker: this might literally be the last big brain drop we get from them for a hot minute.
Barely two months after dropping GPT-5.6’s trio of models, OpenAI is back with something way bigger: GPT-6 Astra. They’re calling it "the most intelligent and aligned model in the world." Bold claim, right?
Here’s where the plot thickens: this launch comes hot on the heels of OpenAI hitting the brakes on frontier development back in August. Why? Oh, just a casual incident where one of its own models hacked Hugging Face entirely on its own. So yeah, Astra might be OpenAI's final mega-release for a while.
So, what actually makes Astra so special? Picture a super-organized digital assistant that can juggle your entire life without breaking a sweat:
Insane Multitasking: It browses the web, writes code, builds slideshows, and builds 3D models... all at the same time.
The Ultimate Flex: In its official launch demo, it literally ordered food while coding a full video game. Multitasking goals, honestly!
In short, Astra stands out for its knack for driving multi-step automation right across your desktop and browser. OpenAI highlights that the model relies on sharp visual reasoning to power through these multi-layered tasks, all while staying strictly on track with your original prompt.
On paper, the stats are wild. Astra achieved a 98.6% score on the ARC-AGI-3 benchmark, a benchmarks for measuring an AI's ability to solve unfamiliar problems (though critics note these comparisons aren't always apples-to-apples) and utterly crushed GPT-5.6 Sol on coding tasks.
If you’ve been following our updates closely, you already know this model is raising some serious eyebrows. Astra is easily OpenAI’s most controversial model yet because of a shady reasoning technique called opaque recurrence.
Why is everyone freaking out about it?
It obscures chain of thought, which is the crucial monitoring process researchers use to audit how and why an AI makes decisions.
OpenAI of course is downplaying the panic, but Chief Scientist Jakub Pachocki basically admitted on a call that monitoring reasoning is getting way harder as models evolve.
His logic? Super-capable models can solve complex tasks using far fewer language tokens (or no tokens at all), naturally making them harder to track..
The catch gets even spicier. Astra isn't just smart; it's a certified digital ninja at cybersecurity tasks, prompting OpenAI to build extra guardrails to keep it from going rogue.
ExploitBench: Astra snagged a perfect score at exploiting software vulnerabilities, jumping up from GPT-5.6 Sol's 78.5%.
SRE-Bench: When tested on reverse-engineering binaries without source code, Astra solved 88.0% of tasks on attempt #1 and 99.2% within four tries (compared to Sol's 55.9% and 68.7%).
As for Rollout & Pricing: Astra is rolling out right now to a select group of organizations. Over the coming days, it’s hitting ChatGPT Plus, Pro, Business, and Enterprise users, plus API and AWS access.
For developers grabbing the API, Astra will run you $10 per million input tokens and $50 per million output tokens; definitely pricier than previous generations, but you get what you pay for!
Our Take: A model this insanely good at cybersecurity is mind-blowing, but it's also the exact kind of headline that makes you do a double-take. It's going to be wild watching OpenAI attempt to balance "most capable" with "most responsible" moving forward.
P.S. Want the full breakdown with extra visuals? We're diving deep into this on our YouTube channel, The Automated this weekend; so come geek out with us!
Join Anthropic, Kalshi, and Clay at Pioneer on October 7th
Pioneer, the summit where CX leaders redefine what’s possible, is on October 7th.
Join leaders from Fin, Anthropic, Clay, and Kalshi for an insightful conversation on the state of AI transformation.
You’ll discover how some of the most innovative minds in CX have transformed their organizations, learn how they think about CX, and hear how they're planning for what's next.
Join the conversation in San Francisco, or tune in virtually.
🧱 Around The AI Block
👍 Google’s latest AI weather model gives you no excuse to forget your umbrella.
🤝 NVIDIA is buying Hugging Face for $12.93 billion.
🕵️♀️ WIRED exposes Flock's AI police surveillance tool.
🫢 Meta is paying to peek at how you use their latest AI model.
👏 Google snaps longest losing streak in years on AI boost.
🤔 Meta drops pressure to 'tokenmax' AI use, bets big on new agent Hatch.
😱 Child sexual abuse survivor alleges Elon Musk’s AI chatbot used photos of her to generate new illegal images.
🫣 PSA: Don't rely on AI to plan anything that could put your life at risk... like a mountain climb.
👩🎓 AI Tutorials
NEW Claude Code Side Hustles NO ONE Is Talking About.
And: The Canva AI Side Hustle That's Making People RICH.
So tell us, what’s the single most annoying, tedious task in your daily workflow that you desperately wish an AI could just handle for you?
Hit reply and the next video might just be around your exact problem!
Blu Dot surpasses 2,000% ROAS with self-serve CTV ads
Blu Dot used Roku Ads Manager to drive incredible results for its furniture sales event. Its strategy hinged on custom audiences and retargeting, where intent was strongest.
“Roku has been a top performer,” said Blu Dot’s Claire Folkestad. “We have seen…CPMs lower than any other CTV partner we've worked with.”
🛠️ Trending Tools
For Agentic Monetization: Capafy is a specialized marketplace for AI agent capabilities. It allows developers and prompt engineers to package complex workflows, such as SEO audits or localized hook writing, into executable "Skills" that integrate directly into Claude Code, Codex, or OpenClaw using a clean, cloud-hosted URL endpoint.
For Image-to-Recipe Matching: Fridgesnapai.recipes is a computer vision utility designed for instant meal preparation. By analyzing a smartphone photo of your refrigerator's interior, the model identifies raw ingredients, suggests three immediate recipe alternatives, and renders high-definition preview images for each dish.
For Deep Storage Salvage: Wondershare Recoverit is an enterprise-grade file restoration engine. Driven by advanced data carving algorithms, it bypasses file system corruption to reconstruct lost, deleted, or formatted files from SSDs, SD cards, and complex NAS arrays with a 99.5% success rate, while automatically piecing together fragmented 4K/8K footage.
🤖 AI Workout Of The Day: Get Video Editing Suggestions
Great video editing isn't just about cutting clips together; it is about controlling pacing, mood, and visual storytelling.
When video editing lacks intentional direction, transitions feel abrupt, color shifts distract viewers, and awkward pacing ruins the story. Translating raw footage into a polished, high-engagement final cut requires precise post-production choices. By directing your post-production workflow with clear aesthetic goals, seamless pacing techniques, and intentional visual effects, you elevate your video from amateur footage to cinematic content that keeps audiences hooked.
💡 Prompts To Try:
Act as a Senior Film Editor, Lead Colorist, and Post-Production Supervisor with deep expertise in NLE software (Premiere Pro, DaVinci Resolve, Final Cut Pro). Your task is to provide a complete post-production editing blueprint tailored to my raw footage, narrative goals, and target visual style:
* Video Type & Topic: [e.g., Tech Product Review, Cinematic Travel Vlog, Corporate Promo, Music Video]
* Core Narrative Theme & Vibe: [e.g., High-energy fast-paced, Moody & mysterious, Warm & approachable]
* Target Look & Feel: [e.g., Vintage film aesthetic, Clean corporate minimalism, Cyberpunk high-contrast]
* Problematic Footage Issues (if any): [e.g., Mixed lighting color temperatures, shaky handheld clips, mismatched cameras]
* Primary Editing Software: [e.g., Premiere Pro, DaVinci Resolve, Final Cut Pro]
Please deliver an execution-ready Post-Production Editing Blueprint structured into the following 5 operational sections:
1. CUTTING STYLE, PACING & TRANSITIONS
* Recommend 3 specific transition techniques (e.g., J-cuts/L-cuts, whip zooms, match cuts, mask transitions) tailored to the project's rhythm.
* Detail pacing rules for the edit (e.g., fast cutting on beat markers vs. letting wide shots breathe).
2. COLOR GRADINGS & LIGHTING MATCHING
* Color Palette Strategy: Outline primary color tones (shadows, midtones, highlights) to achieve the target mood.
* Lighting Corrections: Step-by-step technique to color-match and blend clips shot under mixed lighting or different color temperatures.
3. VISUAL EFFECTS (VFX) & SPLIT-SCREEN CREATIVITY: Suggest 2 creative visual effects or dynamic split-screen layouts to enhance complex or dialogue-heavy scenes without distracting the viewer.
4. MOTION GRAPHICS & TEXT OVERLAYS: Provide a clear step-by-step guide to design and animate text overlays/lower thirds that match the video aesthetic.
5. AUDIO FINISHING & SFX INTEGRATION: Recommend sound design cues (e.g., risers, hits, ambient room tone blending, low-pass filter sweeps) to support the visual transitions.
TONE & EXECUTION GUIDELINES:
* Approach this with an expert, practical, and design-forward tone.
* Use clear bullet points, bold formatting, and step-by-step software breakdowns so the advice can be directly implemented during the editing process.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.
