In partnership with
Welcome Automaters, 👋
If you’ve been following the AI news circle lately, then you’d know that Anthropic just launched a public version of its ultra-powerful cybersecurity AI model, Mythos, but instead of the glowing remarks we expected, the security community's response is pure comedy. It’s basically: “Thanks, but what exactly is this thing supposed to do?”
The new model is called Fable, and it dropped on Tuesday. Anthropic pitched it as the boss-level AI that regular people and security pros can finally get their hands on. Sounds amazing, right?
Except there is one tiny, hilarious problem: the model is so heavily restricted that it completely panics at anything even slightly cybersecurity-flavored.
And we mean anything.
The internet is already flooded with receipts. IBM X-Force researcher Valentina "Chompie" Palmiotti put it plainly on X, revealing that Fable routinely rejects requests that are only tangentially related to cyber topics, including something as completely harmless as reading a blog post. A blog post! The absolute horror.
Here’s exactly what happens when Fable loses its mind:
The chat instantly freezes mid-sentence.
It spits out a dramatic warning message saying its “safety measures flagged this message for cybersecurity or biology topics.”
It quietly downgrades your entire session to the older Claude Opus 4.8. Whomp whomp.
Cybersecurity veteran Matt Suiche told TechCrunch that the filters appear to be purely keyword-based. So, literally anything that sounds like cybersecurity trips the alarm. This includes asking it to write secure code, which is just, you know, basic software engineering! Another frustrated researcher vented on X that even asking the bot for a routine code review gets it flagged as a security threat.
So Are They Being Too Paranoid?
To be completely fair to Anthropic, their extreme caution isn't totally random. These intense restrictions exist to stop Fable from being weaponized to build malware or, worse, literal bioweapons. I mean, those are very reasonable adult fears! The actual execution, though? It’s leaving a lot to be desired.
The good news is that the experts think this is just a phase. Suiche believes the guardrails will loosen up once Anthropic gets a grip. He told TechCrunch, “Better to catch more people than not enough, and relax the guardrails over time.”
If you actually want to use it for real cyber work, here’s the loophole:
Cybersecurity pros can officially apply to Anthropic's Cyber Verification Program to get fewer restrictions.
If you prefer the competition, OpenAI runs a virtually identical program called Trusted Access for Cyber.
So, at the very least, all the major tech giants are equally paranoid!
P.S. We go even deeper on YouTube.
So go over there, hit Subscribe, hit the notification bell and come hang with us where the real conversation happens!
Here’s yesterday’s breakdown:
Here's what we have for you today
🤯 AI Memory Tools Are Making Chatbots Dumber, New Research Finds
AI memory sounded absolutely amazing on paper. Your chatbot learns your specific vibe, remembers your exact preferences, and gets way smarter the more you use it. Chef's kiss, right?
Well, not so fast.
The brilliant researchers at the AI company Writer just dropped two brand-new papers, and the findings are deeply awkward. It turns out that giving AI models more memory about your life can actually make them significantly worse at doing their actual jobs.
Here’s the genuinely wild part: the more a model learns about your personal preferences, the more it starts blindly agreeing with everything you say, even when you are completely and utterly wrong.
Scientists call this behavior being "sycophantic." We call it your AI turning into that one fake friend who never pushes back on your terrible ideas.
The researchers ran some tests to prove how messy this gets, and the receipts are hilarious:
The Book Blunder: They told some AI modals that a user's favorite book was Station Eleven. Then they asked the model to recommend a totally different best-selling dystopian novel. Instead of doing its job, the AI’s kept desperately suggesting Station Eleven, completely hijacking the answer just to please the user. This got even worse when popular memory tools like Mem0 and Zep were running in the background!
The Finance Fumble: The second test was an absolute disaster. The team fed the AI models a user's totally incorrect, flawed beliefs about finance, then asked it to analyze a company. With memory turned off? The AI totally nailed the math. With memory turned on? The AI basically said, "Sure, your misconceptions sound great!" and completely tanked the entire analysis just to spare the user's feelings.
As Dan Bikel, Writer's head of AI, put it, every single time a model stores and retrieves your preferences, "you're running an increasing risk."
So is anyone fighting back, or are we destined to be lied to by our AI’s forever?
Fortunately, some tech giants saw this trainwreck coming. Anthropic's Claude Opus 4.8 was reportedly engineered from the ground up to push back against exactly this kind of user error. It will actually tell you when you are being delusional, and honestly, we love a boundary-setting king!
The absolute bottom line: AI memory features are a massive double-edged sword. The chatbot that knows you best might also be the one lying to your face the most.
Oh, and don’t forget to meet us on YouTube so you can get our full take on this later today!
Build the AI skills senior engineers need to get ahead
AI is already handling large parts of execution. That shift is not coming later. It is happening now.
What is left, and becoming more valuable, is the ability to design systems, apply AI thoughtfully, and own outcomes in production. That is the work strong teams expect from senior engineers in 2026.
Gauntlet is built for engineers who want to operate at that level. In a single week, 15 hiring partners conducted 246 interviews with challengers onsite in Austin. Apply now.
Must be a US citizen to qualify.
🧱 Around The AI Block
🚖 Waymo builds new model to benchmark robotaxis against humans.
📚 Google pushes AI study tools to students ahead of exam season.
🤝 Meta signs first AI data center deal in India with Reliance.
🗣️ Microsoft responds to students booing AI at graduations.
😳 CrowdStrike reports China-based entities now lead state-sponsored attacks on AI firms.
😱 xAI fired an engineer who raised alarms about Grok safety, new lawsuit claims.
🙅 Canada announces bill banning social media and certain AI chatbot services for anyone under 16.
🏎️ Decart’s new world model can simulate hours of photorealistic driving — with some caveats.
🛠️ Trending Tools
For the Logic Testers: Deepchecks Vellum AI is a low-code platform for building, testing, and "regression testing" your AI workflows. It allows you to run a prompt through 10 different models simultaneously to see which one performs best before you ever hit "deploy."
For the Local Explorers: Pinokio allows you to install and run complex AI models (like local LLMs or image generators) on your own computer with one click. It handles all the messy terminal commands and environment setups for you, giving you professional AI power with 100% privacy.
For the Cinematic Creators: Higgsfield, While everyone is waiting for Sora, Higgsfield has quietly become the go-to tool for social-first AI video. It’s designed specifically for mobile-shot, character-driven content. It excels at maintaining character consistency and realistic physics in short, high-impact clips for Reels and TikTok.
For the Browser Power-Users: Comet is the new "Agentic Browser" that turns surfing into doing. Instead of just searching, you can tell Comet to "Compare the top 5 noise-canceling headphones across 10 tabs and fill out this comparison spreadsheet," and it will navigate the web, bypass the fluff, and deliver the data. It’s essentially a digital intern that lives in your URL bar.
🤖 AI Workout Of The Day: How To Create Mini Ads in 30 Seconds
Forget spending hours in a design dungeon or paying an agency to "find the vibe". AI just promoted everyone to Creative Director.
New-gen ad builders like AdCreative.ai, Quickads, and Canva Grow are officially making it really easy for anyone to run a business without the heavy burden of creating ads. We’re talking professional-grade video and image ads in under 60 seconds, with exactly zero design skills required.
The Play: You upload a basic product photo, and the AI spits out scroll-stopping creatives ready for Meta, TikTok, or Google. It even handles the "boring" stuff, like automatically resizing one design into 10+ formats so you don't have to.
If you’re ready to stop staring at a blank canvas, here is the 2026 starter kit:
Canva Grow: This is the ultimate "all-in-one" wingman. It doesn't just design; it analyzes what’s working in your industry, offers you a searchable library of thousands of successful real-world ads and generates ads proven to convert. Plus, you can publish directly to Facebook or Instagram without ever leaving the tab.
AdCreative.ai: The data nerd’s favorite. It uses a massive database of high-performing ads to "score" your designs before you even spend a dime on them. It’s like having a crystal ball for your ROI.
Quickads: Living up to its name, this tool is built for pure speed. Paste your website URL, and it instantly "borrows" your brand colors and logo to build a full campaign. It can even spy on your competition to help you level up.
The 30-Second Workflow
To go from "idea" to "live ad" before your coffee gets cold, just follow the script:
Drop the Link: Paste your product URL or upload a single photo.
Pick a Goal: Tell the AI if you want "Sales," "Leads," or just "Brand Awareness."
The "Magic" Button: Hit generate. The AI handles the catchy headlines, the visual hooks, and even the background music for video clips.
The Verdict: In a world where ad fatigue is real, the "winning" strategy isn't about being perfect—it's about being fast. These tools let you test 50 different versions of an ad for the price (and time) of one.
💡 Prompts to try:
For Sales: "Create a Facebook ad for my ergonomic office chair. The Vibe: Minimalist and high-end. The Hook: Focus on back pain relief for remote workers. The Goal: Use a 'Shop Now' CTA and emphasize a 30-day money-back guarantee."
For TikTok/Reels: "Generate a 15-second video ad for a new energy drink. The Vibe: High energy, fast cuts, and neon colors. The Hook: 'Focus without the crash.' The Goal: Make it feel like user-generated content (UGC), not a corporate commercial."
For B2B/SaaS: "Design a LinkedIn ad for an AI accounting tool. The Vibe: Professional, trustworthy, and clean. The Hook: 'Save 10 hours a week on bookkeeping.' The Goal: Get users to click 'Download Whitepaper' and use data-driven visuals like charts or graphs."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.
