PewDiePie, PewDiePie AI.

@antdaviscyber

22.6K followers · usually 9.9K views a post

Views

130.8K

Likes

3.6K

Comments

170

284

SoundTikTok sound · original sound - antdaviscyber

Why it worked

The video leverages a high-profile name (PewDiePie) to discuss a timely and controversial topic (AI security), creating a sense of urgency and educational value for a broad audience.

Summary

The creator discusses a new open-source AI project called Odysseus, which was promoted by PewDiePie as a local alternative to ChatGPT. He highlights the security risks associated with using AI-coded tools that have broad system permissions, especially when used by non-technical individuals.

Structure

1Introduction to PewDiePie's AI project, Odysseus
2Explanation of the project's local, open-source nature
3Discussion of the 'Vibe Coded' development process
4Analysis of security vulnerabilities found in the tool
5Warning about the risks of AI-assisted tools with system permissions

On-screen text

PEWDIEPIE PEWDIEPIE VIBE CODED AI PEWDIEPIE AI HE JUST TOLD 110 MILLION SUBSCRIBERS TO STOP PAYING FOR CHATGPT AND RUN THEIR OWN AI INSTEAD HE'S GOT A PROJECT IT'S CALLED ODYSSEUS IT'S FREE IT'S OPEN SOURCE 48 HOURS IT HIT 30,000 STARS ON GITHUB NOW GONE PAST 77,000 AND THE PITCH IS REALLY QUITE SIMPLE INSTEAD OF SENDING YOUR CHATS TO OPENAI OR ANTHROPIC YOU RUN THE WHOLE THING ON YOUR OWN COMPUTER THE DATA NEVER LEAVES THE MACHINE WHICH SH- WHICH SHOULD WHICH IS A GOOD THING RIGHT CHAT AI AGENTS EMAIL MANAGEMENT CALENDAR EVEN A RESEARCH TOOL THAT READS THE WEB FOR YOU AND INSTALL IT JUST YET PEWDIEPIE HAS BEEN OPEN ABOUT HOW HE BUILT IT HE CALLS IT VIBE CODED THAT MEANS MOST OF THE CODE WAS WRITTEN BY AI MODELS NOT HUMANS SITTING DOWN AND REVIEWING EVERY SINGLE LINE AND THIS IS A TOOL THAT CAN TOUCH YOUR FILES RUN COMMANDS ON YOUR SHELL OR COMPUTER AND READ YOUR EMAIL OBVIOUSLY SECURITY RESEARCHERS WENT LOOKING ALMOST IMMEDIATELY AND THEY FOUND A VULNERABILITY THAT AN ATTACKER COULD USE THEY COULD EXECUTE CODE REMOTELY RUN CODE REMOTELY THROUGH THE AGENT AND THE FLAW THEY FOUND IT WAS A FLAW IN HOW IT PROCESSED EMAIL THAT COULD BE USED TO SLOW YOUR WHOLE SYSTEM DOWN TO A CRAWL THAT'S JUST ONE THING THEY FOUND THE DEVELOPERS HAVE BEEN PATCHING THESE AS FAST AS THEY'VE FOUND THEM WHICH IS A GREAT SIGN BUT IT DOES TELL YOU SOMETHING IMPORTANT WHEN MILLIONS OF NON-TECHNICAL PEOPLE INSTALL A TOOL WITH PERMISSIONS TO TOUCH THEIR FILES AND THEIR INBOX AND THAT TOOL WAS BUILT PRIMARILY BY AI WITH A SMALL TEAM REVIEWING IT THAT GAP BETWEEN CONVENIENCE AND EXPOSURE IT GETS VERY SMALL VERY FAST AND THIS ISN'T REALLY A STORY ABOUT PEWDIEPIE AND HIS AI ALTHOUGH IT IS BUT IT ISN'T IT'S A STORY MORE ABOUT WHAT HAPPENS WHEN AI-ASSISTED CODING MEETS AI AGENTS WITH REAL PERMISSIONS AT MASSIVE SCALE RELEASED TO AN AUDIENCE WHO HAVE NO IDEA WHAT AN RCE VULNERABILITY IS DO YOU KNOW THIS WON'T LOOK SELF-HOSTING YOUR AI CAN ABSOLUTELY BE SAFER THAN THE CLOUD BUT AND A BIG BUT BUT ONLY IF YOU UNDERSTAND WHAT YOU JUST GAVE IT ACCESS TO

Original caption

PewDiePie’s new AI tool has 77,000 GitHub stars and had a vulnerability that let attackers run code remotely. #cybersecurity #ai #pewdiepie #datasecurity #techtok

More from @antdaviscyber

More breakout videos from this creator.

Transcript

PewDiePie, PewDiePie AI. He just told 110 million subscribers to stop paying for ChatGPT and run their own AI instead. He's got a project, it's called Odysseus. It's free, it's open source, and in the first 48 hours, it hit 30,000 stars on GitHub, and that number's now gone past 77,000. And the pitch is really quite simple. Instead of sending your chats to OpenAI or Anthropic, you run the whole thing on your own computer. The data never leaves the machine, which should, which is a good thing, right? You know, chat, AI agents, email management, calendar, even a research tool that reads the web for you. Don't run off and install it just yet, hang on. PewDiePie has been open about how he built it. He calls it Vibe Coded. That means most of the code was written by AI models, not humans sitting down and reviewing every single line. And this is a tool that can touch your files, run commands on your shell or computer, and read your email. Obviously, security researchers went looking almost immediately, and they found a vulnerability that an attacker could use. They could execute code remotely, run code remotely through the agent. And the flaw they found, it was a flaw in how it processed email that could be used to slow your whole system down to a crawl. That's just one thing they found. The developers have been patching these as fast as they've found them, which is a great sign. But it does tell you something important. When millions of non-technical people install a tool with permissions to touch their files and their inbox, and that tool was built primarily by AI with a small team reviewing it, that gap between convenience and exposure, it gets very small, very fast. And this isn't really a story about PewDiePie and his AI, although it is, but it isn't. It's a story more about what happens when AI-assisted coding meets AI agents with real permissions at massive scale, released to an audience who have no idea what an RCE vulnerability is. I mean, do you know? Many people watching this won't. Look, self-hosting your AI can absolutely be safer than the cloud. But, and a big but, but only if you understand what you just gave it access to.