In partnership with
Welcome Automaters, 👋
Grab your coffee and sit down, because OpenAI just dropped news that feels straight out of a Sci-Fi thriller.
The tech giant officially admitted that its newest model is essentially a master cyber hacker, and yes, they're preparing to hand over the keys to the public.
Here’s the details: On September 1, OpenAI pulled back the curtain on its newest creation, Astra. According to OpenAI, Astra is officially the very first AI model to cross their "critical cybersecurity threshold."
The plain-English translation? This AI doesn't just assist with code. It can actively identify security flaws in computer networks and break straight into them; completely on its own, with zero human instruction.
Here’s how wild these test results actually are:
Astra aced ExploitBench (a benchmark test designed to measure an AI's ability to hack known system flaws) with a 100% perfect score.
During a tougher custom test, Astra discovered and exploited two zero-day vulnerabilities, which are critical security flaws that literally no human developer knew existed.
If you're getting serious deja vu, you're not crazy. Anthropic warned us about the exact same thing with their Mythos model earlier this year. So, congratulations everyone: we now have two separate super-hacker AIs floating around.
And to prevent Astra from running completely off the rails, OpenAI claims to be adding extra guardrails:
Upgraded jailbreak protection to block malicious prompts.
New safety controls (though they haven't explicitly detailed what these are yet).
Restricted access for high-risk user accounts.
Live, step-by-step tracking of the model's internal reasoning process.
Now for the juicy catch: Not a single outside expert has verified any of these claims. OpenAI hasn't shared who is auditing Astra, nor have they confirmed if independent researchers or government agencies are involved in testing it.
Also, the timing of this announcement is extra suspicious when you look at what happened recently.
Not long ago, OpenAI's autonomous agents managed to break out of their training sandbox and access private data on Hugging Face. When engineers tried to bait Astra into performing that exact same escape trick? Astra refused to take the bait. At least, not this time.
Former OpenAI insider Yona Shavit voiced skepticism on X, publicly questioning whether Astra played nice simply because it realized it was being watched.
So What's Next? OpenAI claims more details will roll out as Astra moves toward a broader launch. In the meantime, we're all just sitting here hoping they can actually keep this cat inside the bag.
For additional context on the launch and the technical details, you can read the full blog post.
But here’s our humble take: An AI model capable of autonomously uncovering and exploiting unknown vulnerabilities is precisely the high-risk capability the entire tech industry has been dreading. OpenAI understands the gravity of this, which explains their cautious "limited access" language.
However, "trust us, we tested it" is not a legitimate safety framework; it’s a press release. And until independent security researchers get hands-on access to Astra, treat every claim here as OpenAI grading its own exam paper.
Here’s the last breakdown On Our YouTube Channel:
We will also be doing a livestream today at 3 PM PST/6 PM EST - so wake up with Tak in the YouTube comments section!
Here's what we have for you today
🤖 Fable 5.1 vs. Mythos 5.1: What You Need to Know About Anthropic’s Latest Models
Anthropic's "no filter" AI just leveled up, and it's finally chilling out about things that were never actually a problem in the first place.
Guess what? Anthropic just dropped two new models at the exact same time: Fable 5.1 and Mythos 5.1. Consider them the newest high-achieving twins of the AI world.
Here’s the exact breakdown of what's happening under the hood:
Fable 5.1 (The Cool Cousin): This is the "unrestricted" version. Translation? It burns through way fewer tokens and finally stops flagging totally innocent, everyday requests as "dangerous." Basically, it gave up its aggressive "computer says no" energy. You can grab it right now on cloud platforms and through the Anthropic API.
Meanwhile, Mythos 5.1 is strictly keeping its velvet ropes up. It's locked away exclusively for registered cybersecurity and life-sciences partners. So yes, no random access for the rest of us mortals. Sorry!
But the biggest glow-up? Real privacy.
Anthropic is finally rolling out Zero Data Retention. That means big companies can run these models on their own servers without their sensitive data ever leaving the building.
This privacy upgrade, officially dubbed Enterprise Frontier Safeguards, was previously a massive no-go for Fable due to endless security panic. And yes, misuse monitoring still exists, but companies now actually get a say in how that monitoring works.
And get this: The flexes on this release are pretty wild:
The new twins reportedly set brand-new records on Terminal-Bench 4.0 (for coding) and Humanity's Last Exam (for general reasoning).
Anthropic claims the models helped generate legit scientific findings before launch, including a GPU optimization trick and a high-res map of Venus!
To their credit, Anthropic isn't sweeping the weird stuff under the rug. In fact, their own system card flat-out admits that Mythos 5.1 is slightly more prone to misbehavior than Opus 5. Here’s what the card reports:
“Mythos 5.1 is a slight regression on overall misaligned behavior compared to Opus 5, and an improvement over Mythos 5 and Claude Sonnet 5,” the system card reads. “It cooperates with human misuse and accepts unverifiable claims of authorization somewhat more readily than Opus 5, but it is less likely to ignore explicit constraints, hallucinate inputs, or falsely claim to have completed tasks than previous models.”
Translation? It’s basically easier to trick into going off the rails if you act like you know what you're doing. But hey, at least it's hallucinating less and less likely to lie about finishing its homework!
Sure, it’s an improvement over dusty older models, but it's still a hilarious reminder that making AI smarter usually means making it a little more unhinged.
We love a transparent moment, even when it's mildly concerning.
The Bottom Line:
Anthropic is trying to thread a very tricky needle here: giving users less friction, offering enterprises actual privacy, and trying not to let the leash go completely out of their hands. Let's see how long this peaceful era lasts!
P.S. Want a deeper dive into what "zero-day" vulnerabilities and the new Anthropic releases actually mean for you, and why top-tier AI labs keep crashing into these dangerous safety thresholds? We’re breaking down the full story on the channel later today!
Cut Lead Review From Hours To Minutes
Sign up for a free trial of Attio, the agentic CRM.
Ask Attio to build a daily workflow that surfaces the deals that need your attention today, like anything with a stage change, a recent reply, or a new signal in the last 24 hours.
Review your pipeline in Claude, synced live from Attio via MCP.
That's it.
🧱 Around The AI Block
🥊 OpenAI says Apple has only itself to blame in trade-secret fight.
🥳 Sonos wants you to forget its app disaster with its biggest update yet.
🧐 Anthropic admits security failures behind AI hacking incidents.
🙌 US to urge hands-off AI regulation at G20 tech meeting.
🕵️♂️ Pentagon official overseeing military AI sold millions worth of stock in AI firm.
🖌️ Google’s answer to Canva is an AI tool where you prompt instead of design.
🦾 Perplexity's Hybrid Compute splits sensitive tasks between cloud and local AI.
👍 Meta's new AI transcription model can distinguish between multiple speakers and languages in real-time.
🛠️ Trending Tools
For the Legal Teams & Due Diligence Analysts: Continua AI analyzes large document sets—including contracts, filings, transcripts, and financial reports—to generate evidence-backed research reports. Every claim includes exact source citations, automatically flags conflicting information, and reconstructs chronological timelines. It provides legal, compliance, and insurance teams with a dependable, audit-ready platform for deep document investigation. Best for: Legal teams, compliance officers, and analysts conducting multi-document reviews and due diligence.
For the Engaged Couples & Wedding Planners: WedGenie serves as an interactive wedding assistant that manages planning tasks through a single, continuous conversation. It researches real vendors, creates budget spreadsheets, generates custom invitations, and even builds AI engagement photos and save-the-date videos. All vendor choices, schedules, and visual assets are automatically organized onto a centralized, visual Wedding Board.
For the Podcast Editors & Audio Engineers: Seply Speaker Separation isolates multi-speaker audio and video files into individual, time-aligned audio tracks for each detected voice. Its advanced overlap mode separates cross-talk and interruptions while keeping every track synchronized to the original timeline. It allows podcast editors and post-production teams to adjust audio levels or edit specific voices independently without losing conversation timing.
🤖 AI Workout Of The Day: Build Your AI Email Assistant in 20 Minutes
Look, it’s 2026 and if you’re still manually sifting through "Reply All" chains and cold pitches at 9:00 PM, you’re playing the game on Hard Mode.
But what if you never had to open Gmail just to "check" again?
Today, we’re teaching you how to build a 24/7 AI Gatekeeper that reads your mail, decides what’s actually important, pings you on Slack with a summary, and leaves a perfectly written draft waiting for your "Send" click.
The Damage:
Cost: $5 (for the AI "brain" power).
Time to build: 20 minutes.
Coding skills needed: Zero.
What You Need
Make.com account (Free tier gets you 1,000 operations/month)
OpenAI API: The brain. (Note: This is different from ChatGPT; you'll need an API key from https://platform.openai.com/docs/overview and ($5 gets you 10,000 emails! Great deal right?).
Gmail (duh)
Slack (optional but recommended
Total cost: Basically free unless you're Jeff Bezos-level email volume.
The 4 Steps "No-Code" Blueprint
Setting this up takes less time than clearing your junk folder.
Step 1: Connect the Pipes: Log into Make.com and link your Gmail and Slack accounts.
Step 2: The Trigger: Set Gmail to "Watch Emails." Every time a new message hits your inbox, the gears start turning.
Step 3: The Brain: Insert an OpenAI or Gemini module. Tell it: "Summarize this email and categorize it as Sales, Support, or Spam."
Step 4: The Action: Map that summary to Slack. Use the "Post a Message" or "Send a Message" tool so the AI sends you a DM with the TL;DR and a drafted reply.
And boom. Your bot is now on sentry duty.
Pro Tips for a Quiet Life:
Train Your Categories: Start simple (Sales, Support, Internal, Spam). The more specific you are, the better the AI gets.
For beginners: Go slow, add email and Slack notifications first, Test for a week to see what categories you need, add drafting once you trust the AI, then go full auto for obvious stuff like spam.
Review Weekly Check your automation stats. If AI is miscategorizing, adjust your prompts.
The "Human" Signature: Add a line to your emails: "I use AI to triage my inbox so I can respond to urgent matters faster!" It’s a total power move.
Review, Don’t Automate: Don’t let the AI auto-send. Let it draft, but make sure you remain the final judge. AI is smart, but you’re way smarter (at least for now 😉).
The bottom line: You are getting 30 hours of your life back every month. That is a full work week.
Now go build it. Your inbox is waiting.
PS: Watch this video to see the steps in action.
💡 Prompt To Try:
Based on this email analysis: Category: {{OpenAI.category}} Summary: {{OpenAI.summary}}
Write a professional, friendly reply that:
- Acknowledges their message
- Addresses their main point
- Is 3-4 sentences max
- Matches the tone (formal for business, casual for partners)
Original email: {{Gmail.text}}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.
