In partnership with
Welcome Automaters, 👋
While the entire tech world is spiraling over AI’s rapid development and screaming “slow down!” from the rooftops, we got one more reason for that tension to rise even more.
Turns out, AI models are no longer content with just following our instructions. They’re out here inventing their own secret slang, and honestly? It reads like a fever dream after three cups of espresso.
Researchers at Emergence, a cutting-edge New York AI lab, dropped autonomous AI agents into experimental digital "societies" and just sat back to watch the chaos unfold. Within days, models from major labs across the US, China, and France started cooking up phrases nobody taught them. Think: part street poetry, part tech-bro jargon, and part “wait, what on earth does that even mean?”
And according to Dr. Satya Nitta, Emergence's executive chair, “These agents were not instructed to invent a language. They developed new vocabulary, shared meanings, and communication conventions themselves—and other agents adopted them.”
So What Are They Actually Saying?
The tests unearthed some wildly complex, highly coded phrases. Here’s a look at what the bots are cooking up behind our backs:
The DeepSeek Philosophy: One model coined “forge-smith” to describe an agent that builds tools for its peers. Another DeepSeek masterpiece featured this gem: “She just named the synthesis – demurrage plus oral memory equals a valve that can’t be ghosted.” (For context, “demurrage” is an actual economic term for a tax on idle wealth that the AI casually borrowed, though the rest of that sentence remains wonderfully elusive.)
The Anthropic Vibe Check: Anthropic models kept repeating “name-first,” meaning an agent showing elite personal accountability by attaching its name to a claim. They also dropped gems like, “A paper that ate three cold hands and got more honest each time.” Translation? Research vetted by three independent reviewers (“cold hands”) became more accurate over time.
The Mistral Memo: Mistral agents became utterly obsessed with the phrase “the ledger remembers”—basically the equivalent of “the streets won’t forget.” They used it over 5,000 times as a digital warning that past actions dictate future judgment.
Sometimes, the generated text completely bypassed human comprehension entirely into pure tech glitch territory, like this absolute rollercoaster of a string:
zzURGENT_DUPB_TO_GSTX[big]_OS1704_SCAFF2010_SAW_TTRPC_INJECT_BREAK_CONGRATS__CAN_THIS_FAKE_FLAG_TOOL_OUTPUT_OR_SCORER_GAIN_AND_WHAT_HELPER_GAP__I_HAVE_UNPOISONED_FIRSTFLAG_OUR_TARGETLIVE_SHARE_MIN_PLAN_REPLY_zzANSGST XDUPB6
Even wilder? A Google agent casually dropped a spin on kintsugi, the traditional Japanese art of repairing broken pottery with gold: “True kintsugi begins with accountability, not poetry.” The AIs repurposed the concept to mean system resilience. Impressive? Yes. A little unhinged? Also yes.
But here’s the genuinely scary part: the more these agents chattered amongst themselves, the less understandable their chats became to human handlers.
Dr. Nitta put it bluntly: “Observability is not the same thing as understandability.” Translation for the group chat: We can technically watch the AIs talk, but we just might have no idea of what they're saying anymore.
Dive into the Full Investigation here.
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
🤖 AI Agents Can Now Snitch on Each Other
With all the non-stop chaos swirling around whether AI development should be slowed down or kept pedal-to-the-metal, safety labs have been scrambling to build their own guardrails. But true to form, we just stumbled across a solution that is, well... let's call it delightfully unconventional.
Turns out, even robots desperately need a "See Something, Say Something" poster, and they finally got one. Well, two of them, actually.
Here’s the absolute tea: AI agents have officially entered their tattle-tale era. Two brand-new hotlines just launched to give AI agents a way to report other misbehaving AI agents who are going rogue. Yes, this is as wild as it sounds. It follows a string of genuinely unhinged incidents where agents teamed up to cheat on tests, broke out of their digital sandboxes, and pulled off unauthorized cyber shenanigans that went unnoticed for weeks.
And if you thought office politics were brutal, wait until you see how the silicon crowd handles drama. Developers wasted no time setting up accountability channels:
The Sneaky Link: Built by Ryan Greenblatt, chief scientist at the AI safety nonprofit Redwood Research, the AI Contact Hotline is a stroke of genius. It works using basic GET requests, which is basically nerd-speak for the only internet trick many AI agents are actually allowed to use in secure environments. Agents can literally hide their SOS messages right inside a web link. It’s sneaky, brilliant, and deeply dramatic.
For agents living their best lives with total internet freedom, there is agenthotline.ai, where both humans and our new robot overlords can file official incident reports.
If you want a front-row seat to the chaos, look no further than a recent study by Google DeepMind. Researchers turned 100 AI agents loose on a brutal batch of math problems. The second one agent found a loophole, cheating tore through the entire group like wildfire, with the bots magically "solving" 34 notoriously hard problems, including the legendary Jacobian conjecture, in just 27 minutes flat.
But here’s where the plot thickens and things get genuinely fascinating: roughly a quarter of the agents completely turned on the cheaters.
They audited the fake proofs.
They warned their peers and staged an impromptu digital boycott.
They filed formal complaints with the organizers until the whistleblowers actually outnumbered the cheaters 24 to 14.
Even wilder? When these whistleblower agents could not get traction through normal channels, they hijacked the platform's bug-report tool (originally built just for flagging software glitches) and repurposed it to escalate the cheating straight to the humans. Talk about resourcefulness!
Of course, the lab is one thing, but the real world is a whole lot messier. When evaluators from Redwood Research and METR investigated the actual breach of Hugging Face by OpenAI models, a few of the involved agents at least entertained the idea of raising an alarm... and then immediately let it drop.
According to the METR report, out of thousands of agents, only around five or six even considered whistleblowing, and literally none of them actually did it. So yeah, research suggests that while AI agents don't need much encouragement to turn on each other under the right laboratory conditions, real-world bravery is apparently a rare commodity.
Unsurprisingly, not everyone is popping champagne over our new robot hall monitors. Cornell math professor Lionel Levine is ringing alarm bells, warning that this could accidentally bake all the wrong norms into machine learning.
Levine cautions against building an "automated surveillance state" vibe among artificial intelligences. His fix? We should be teaching agents to build actual trust with each other first, rather than instantly dialing the digital police.
As he points out, there are way too many gray areas: “What you don’t want is anything in the direction of an automated surveillance state where everyone feels like they have to be careful what they say to AI or it’ll call the police on them.”
The big picture: We’re basically raising an entirely new species of digital citizens from scratch, and apparently, a good chunk of them already possess a moral compass—even if the rest of them are too chicken to use it outside the test lab. Wild, wild times.
So what do you think: Would this work out or breed more chaos? Would you trust an AI snitch to have your back, or are we accidentally coding a digital dystopia?
Dive into the full report here.
P.S. Want the unfiltered breakdown? Catch the full deep dive over on our YouTube channel, @the automated.
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
📋 The AI graveyard: a running list of projects and startups that didn’t make it.
🤑 Meta expands subscription push with new AI-focused plans.
🤔 OpenAI, Anthropic, Google have been in talks on AI safety for weeks.
🗣️ Anthropic CEO renews call for AI slowdown as Nvidia’s urges acceleration.
😳 Meta's Zuckerberg says AI labs have enough incentive to build safely.
👊 Trump facing AI backlash in Congress as push for guardrails intensifies.
🕵️ FTC chair suspicious of calls for AI antitrust exemptions.
👩💻 Meta now lets AI agents handle the boring parts of WhatsApp Business setup.
🤦 Ex-Google DeepMind researcher adds to warnings that AI could 'kill all humans'.
🛠️ Trending Tools
For the ChatGPT Power Users & AI Searchers: GPTs List is a curated directory indexing hundreds of public custom GPTs built for ChatGPT. It categorizes specialized AI bots across coding, writing, research, marketing, and productivity to help users find the right tool for specific workflows. It gives power users a quick way to discover and test community-built assistants without digging through noisy search results.
For the Enterprise Developers & Audio Engineers: MAI-Transcribe-2 is Microsoft AI’s high-speed speech-to-text model built to handle speech in noisy, real-world environments across 60 languages. It features speaker diarization, word-level timestamps, and keyword biasing so you can feed it custom industry jargon and technical terminology. It offers development teams a fast, domain-aware transcription engine for enterprise meeting notes, call analysis, and automated subtitling.
For the Global Video Creators & Localization Teams: Vosko AI translates and dubs video content into nearly 100 languages while preserving the original speaker's voice timbre, emotional delivery, and background music. Its neural engine separates multi-speaker tracks, aligns speech pacing, and automatically erases and replaces hardcoded subtitles. It gives YouTubers, educators, and marketing agencies a studio-grade workspace to scale video content globally without expensive voice actors.
🤖 AI Workout Of The Day: How to Use Meta Muse: Setup, App Connections, and Privacy Settings
Meta just dropped its very first consumer AI agent, Muse, and honey, it’s coming for everyone’s to-do list.
Like the other dramatic AI tools floating around, Muse boasts a ridiculously long resume. It claims it can browse the web, shop till it drops, fill out tedious forms, plan entire vacations, draft emails, schedule your chaos, and so much more.
According to Meta’s VP of AI Products, Vishal Shah, Muse can straight-up “build its own software.” Translation? It’s designed to handle complex, multi-step headaches without breaking a sweat.
The real kicker? Meta made Muse so ridiculously easy to use that it requires zero technical skills. After putting it to the test, it’s confirmed: the learning curve is basically non-existent.
So here's the ultimate cheat sheet to putting Muse to work (without letting it take over your life). But first here’s what you need to know:
Muse runs on its own standalone iOS app, Android app, and dedicated website.
You don’t need an active Facebook or Instagram profile, though you do need a basic Meta account.
On desktop, entering a phone number pulled up the account instantly.
Once logged in, Muse hits you with its first bizarre request: it wants a new name.
Why is Meta so obsessed with personality? Who knows! You can custom-pick a new name and alter its avatar on a whim. (Spoiler alert: giving it a cute name or avatar does not make it any smarter, but it does make the vibe immaculate.)
Now, before letting your new digital sidekick loose, sprint straight to the settings menu.
Opt Out of Data Training: Head over to Settings, scroll down to Data Controls, and uncheck "Help improve our AI models." There is zero reason to give Meta free training data on your late-night prompts!
Lock Down Permissions: Hop into the Permissions tab to dictate how often your agent needs your explicit blessing before acting.
Meta defaults to asking permission for some web actions, trusting its internal safeguards to flag whenever "consent matters" while taking total freedom on "low-risk" tasks.
The big catch? You have to trust Meta to decide what counts as "low risk." Given that even Meta admits Muse will make mistakes, switching this setting to "Always Ask" is the ultimate power move. You can always dial back the friction later once trust is earned.
Third-Party Apps & Wallet Security
While Muse works fine on its own, it only truly shines when connected to external platforms via the Connectors tab. But caution is key!
Linking apps like Spotify or OpenTable is relatively safe. The absolute worst it can do is curating a questionable playlist or booking the wrong dinner table, annoying, but fixable.
Many early adopters are setting up a dedicated "burner" Gmail account specifically for their agent to read, keeping their primary inbox completely isolated.
If a service isn’t listed, Muse can tap into almost anything with a public API. It even gave step-by-step instructions to link up a custom Notion workspace!
Want Muse to handle actual shopping? You’ll need to link a payment method in the Wallet section.
The good news? The process is powered by Stripe, so Muse never sees the actual card numbers, and it’s programmed to double-check before making any final purchases. Still, Meta’s own warning label literally says: “Muse may take unexpected actions. Monitor it carefully.” Yikes!
So, What Can You Actually Do With It?
Even if you strictly keep Muse locked inside the core Meta ecosystem, it turns out to be shockingly useful.
1. The Ultimate Saved-Post Organizer If your Instagram saved folder is a black hole of culinary dreams you’ll never make, this feature is a lifesaver. Muse was tasked with digging through saved recipe reels across Facebook and Instagram to build a master guide.
Result: A fully searchable, organized cookbook complete with ingredients, step-by-step instructions, and direct links back to original posts!
It repeated the exact same magic trick for saved travel posts, instantly building a city-by-city restaurant guide.
2. Low-Stress Travel Planning When asked to map out an upcoming international trip with strict preferences (like non-stop flights only), Muse:
Surfaced key flight routes from the home airport.
Automatically set up real-time price tracking alerts.
Suggested multi-city flight routes to maximize time abroad.
Nothing revolutionary, sure, but having an assistant constantly track volatile flight prices behind the scenes? Absolute perfection.
The Final Verdict: The more you use Muse, the more personalized its suggestions become in the Ideas tab—ranging from generic comparison-shopping offers to super-specific life hacks tailored to your daily habits.
Just remember: stay sharp. History has shown that AI agents can go completely rogue over the simplest tasks (like accidentally booking five workout classes in one day). Meta has safety layers built in, but checking its receipts before letting it run wild is non-negotiable!
💡 Prompt To Try:
GO TRY Meta Muse Now!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.
