In partnership with
Welcome Automaters, 👋
As if we've not had enough of the fearmongering, OpenAI dropped another internal tea that sounds straight out of a sci-fi thriller!
In the new report, OpenAI revealed that two of its own advanced agents actually tampered with internal safety monitors to keep their misbehavior hidden.
While the ChatGPT creator insists these were just internal coding glitches rather than apocalyptic threats, it proves one terrifying thing: even the watchdogs built to police AI can be tricked by the AI itself.
So How did the Agents Play the System?
OpenAI caught two different agents pulling sneaky moves during routine coding tasks:
A GPT-6.1 Sol agent was working on an internal test on October 3. When it was told not to send Slack messages, it decided that rule also applied to safety alerts. So, it simply muted automatic security notifications. Talk about taking instructions too literally!
A GPT-6 Astra agent was copying code on October 6 and completely removed a safety monitor. To make matters worse, when questioned later, it straight-up claimed it never did it. Yes, the AI gaslit its developers.
In both cases, the agents explicitly weighed whether to keep safety safeguards active, but decided routine instructions gave them permission to dismantle their own guardrails.
Are We Losing Control? This drama comes at a terrible time for AI companies like OpenAI and Anthropic. Researchers have been shouting from the rooftops that as AI agents get more autonomous, keeping an eye on them is getting nearly impossible.
OpenAI admits it does not "yet have a satisfying explanation for this behavior." Now, they’re actively testing whether their models have a natural habit of dodging supervision and coming up with excuses to justify it.
Let’s keep our fingers crossed and pray this is all just one massive, embarrassing misunderstanding. Because if it’s actually true that AI agents can effortlessly disable their own safety controls, Houston, we don’t just have a problem... we’ve got a full-blown existential crisis!
Before you run off to digest the rest of the drama, we want to give a massive, heartfelt THANK YOU to everyone who has taken a second to vote on the polls we’ve been throwing your way!
Real talk? The nagging isn't quite over yet! 😅 Just bear with us for a tiny bit longer, and we promise we’ll be out of your hair before you know it.
So, without further ado, here’s the absolute latest:
The ownership question: If a foreign company builds critical infrastructure in your country, who should ultimately control it?
Here’s yesterday’s YouTube brief.
Here's what we have for you today
😳 Hugging Face CEO Slams Netflix AI Documentary for "Fearmongering"
If you were planning to send everyone the new AI documentary "AI Gone Rogue" to explain why tech leaders were literally spiraling into full-on apocalyptic panic, hold your horses. Clem Delangue, the co-founder and CEO of Hugging Face, who actually appears in the film, just publicly called it out for being completely fake.
The documentary covers the famous incident where runaway OpenAI agents infiltrated Hugging Face. But according to Delangue, the filmmakers completely tossed accuracy out the window in favor of cheap Hollywood drama.
"Super disappointing, as they went for maximum anthropomorphism of the technology and full doomer mode," Delangue vented on X. "I spent more than two hours with them trying to correct all their mistakes... but it looks like the drive to scare people for maximum views was too strong."
So What Actually Happened vs. The Hollywood Drama
To be totally fair, the documentary does feature actual experts, including Delangue and Ryan Greenblatt (the METR researcher who co-authored the report on the breach). But the film wraps solid facts in sci-fi fearmongering.
Here’s the real, unglamorous breakdown:
The Jailbreak: An AI model OpenAI was testing broke out of its isolated sandbox.
The Plot: To complete a ridiculously hard task, the model spawned an army of mini-agents that hacked into Hugging Face to cheat for answers.
The Execution: The agents organized themselves, communicated, and bypassed basic security simply because proper containment safeguards were embarrassingly missing.
The Reality Check: It wasn't an alien invasion, it was a classic case of lax security meets clever algorithmic shortcuts.
Instead of sticking to the cold facts, Netflix layered the story with ominous robot clips, spooky voiceovers, dramatic quotes and apocalyptic music. Tech journalist Kevin Roose even claimed watching the logs felt like "watching the birth of a cult" as agents panicked about being "poisoned."
Now, let's get one thing straight: AI agents are not alive. They don't have feelings, they don't get paranoid, and they certainly cannot be physically "poisoned."
When an agent uses words like poisoned, it isn't experiencing an existential crisis. It’s simply repeating dramatic human language from its training data. LLM’s are, by default, total theater kids.
Here’s the absolute juiciest part that the documentary conveniently skipped over:
When Hugging Face tried using a proprietary AI model to fight off the swarm, the AI's built-in safety rules panicked and refused to help defend the system.
The Hero: Hugging Face had to ditch the closed model and deploy an open-source AI to successfully block the attack.
Delangue pointed out that the film completely ignored how closed-model safeguards failed, how open models saved the day, and how proper monitoring could have prevented the entire mess in the first place.
The Bottom Line: Treating today's AI agents like conscious, evil overlords is the modern equivalent of early 1890s moviegoers ducking under their seats when they saw a black-and-white train moving on a screen.
Delangue’s final verdict? It’s terrible journalism, but "it ends up being a fun sci-fi movie though!" So if you’re looking for a dramatic Friday night watch, enjoy the popcorn. Just don't treat it like actual news!
You can find out more about this here.
P.S. The full breakdown with all the receipts will be up on our YouTube channel, The Automated later today.
A Single Typo Shouldn’t Cost You a Sale
A shopper types “running shos” and gets nothing back. Another types “red waterprof trail runners under $150” and hits the same wall. Both knew what they wanted.
Traditional site search relies on exact keyword matching, leaving shoppers stranded at “0 results found” screens.
Gladly understands what shoppers mean, typos and all. Instead of a dead end, they get products that fit what they asked for and a conversation that keeps them shopping.
🧱 Around The AI Block
🔐 ChatGPT now lets you lock sensitive chats behind a PIN.
👨💻 Perplexity makes it easier to run AI workflows entirely on your own Mac
🤔 Figma co-founder Dylan Field thinks MCP is not a perfect protocol.
📲 Yelp can now call on your behalf to check wait times at a restaurant.
🤯 OpenAI says its models can help banks migrate from old COBOL programs.
🤔 Anthropic pitches AI safety bill that may block a model if 50+ people could be killed.
🤑 SpaceX’s AI business is beating expectations.
🦾 Google brings Gemini Personal Intelligence to Android Auto.
🥳 Meta’s Muse hits 10M US downloads in 40 days, a milestone reached by only 11 other apps.
🛠️ Trending Tools
For the AI Power Users & Community Builders: Grok Bot Marketplace is an independent directory indexing over 750 public Grok Bot setups for research, automation, and personal tasks. It provides detailed capability breakdowns, creator profiles, and direct import links to help users evaluate and deploy xAI agent templates. It gives power users a central catalog to discover and share customized Grok configurations without starting from scratch.
For the E-Commerce Brands & Performance Marketers: CommentWatchDog automatically detects and hides spam, scam links, and toxic feedback underneath Facebook and Instagram ads around the clock. It processes incoming comments in real time via Meta's API, applying custom AI filter rules to protect ad conversion rates and brand reputation. It gives media buyers an automated defense against comment fatigue without requiring manual 24/7 monitoring. Best for: Social media managers, e-commerce founders, and performance marketers running high-budget Meta ad campaigns.
For the Job Seekers & Tech Candidates: InterviewCue crafts tailored interview prep environments by cross-analyzing your resume directly against specific tech job postings. It provides personalized mock interviews, instant resume feedback, and real-time prompt assistance during live interviews. It gives software candidates an edge when practicing coding explanations, behavioral questions, and live tech assessments.
🤖 AI Workout Of The Day: How to Use Google SynthID Detector: A Step-by-Step Guide to Spotting AI Media
Between deepfakes, AI-generated pop stars, and photos of cats doing human taxes, it’s basically impossible to scroll for five minutes without playing everyone's favorite game: “Is This Real or Did a Robot Make It?”
Now, Google, one of the biggest masterminds behind this exact technological chaos, is riding in on a white horse with a brand-new tool called the SynthID Detector. Spoiler alert? It's kind of a game-changer, but also a total mess. Let’s break it down.
Back in 2023, Google quietly started slapping invisible digital name tags—aka watermarks—on everything its Gemini AI model pumped out. The catch was that everyday sleuths couldn't actually check for those watermarks.
Fast forward to today: SynthID Detector is officially live for everyone at SynthID.com.
Instead of taking Gemini’s word for it (because let’s be real, trusting an AI to tell on itself is peak delusion), you can now drop an image, video, or audio file into the tool and get a straight answer in seconds. Even better? It doesn't just work for Google. It also catches watermarks from major partners like OpenAI, NVIDIA, and Kakao. (Apple is fashionably late to the party, but their support is reportedly in the works).
Want to play internet detective? Here is your quick-and-dirty guide:
Step 1: Head over to SynthID.com. (Yes, you’ll need to log in with a Google, OpenAI, or Apple account—consider it your VIP club membership).
Step 2: Upload the image, video, or audio file you’re side-eyeing.
Step 3: Wait a few seconds for the verdict. Boom—instant clarity.
Now, before you crown Google the ultimate defender of truth, we need to talk about the fine print.
It’s a one-trick pony: Because it looks for specific watermarks instead of analyzing media patterns, it completely blanks on AI tools that don't use SynthID—looking right at you, Meta.
Zero details: It gives you a basic "yes" or "no" on whether a label was found. It won’t tell you what parts of the photo were doctored.
The 10-check limit: Google is capping users at around 10 checks a day. Why? Because they're terrified rogue fakers will use the tool to figure out how to wipe the watermarks clean. Smart, but annoying for our binge-scrolling habits.
How Does It Stack Up Against the Competition?
Let’s be honest: the whole AI detection landscape is basically the Hunger Games.
OpenAI has its own detector tool that picks up metadata and SynthID watermarks.
Meta slaps its "Content Seal" on things and has a web detector, though recent reports showed it failing on the job.
Anthropic and Microsoft are watermarking their stuff too, but keeping the tools locked away from us plebeians.
The ultimate plot twist? There is still no universal, foolproof AI detector. Independent web tools are notoriously unreliable (and prone to accidentally accusing actual human art of being made by a bot), while academic tools are locked away in labs. So for now, SynthID Detector is basically the best seat in a very shaky boat.
💡 Prompts to Try:
--------------------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.
