Concerned about where AI is going? Tell your representatives! Go to https://controlai.org/infographics-show to contact your lawmakers in less than 60 seconds using ControlAI’s easy-to-use tools.
Thanks to ControlAI for sponsoring this video! ControlAI is a nonprofit campaigning to ban the development of superintelligent AI internationally to prevent the extinction risk it poses and keep humanity in control.
The opinions in this video are our own, and we're excited to partner with ControlAI on this video to help grow the campaign to prevent catastrophic risks from AI.
AI safety experts spent years worrying about rogue superintelligence, but one of the most dangerous AI security failures came from something far simpler. A poisoned file.
Hugging Face has become one of the most important repositories in artificial intelligence, hosting millions of models, datasets, and applications used by developers around the world.
But some machine-learning models are distributed using Python’s Pickle format, and Pickle files can execute hidden code when they are opened. That creates a massive supply-chain risk.
In 2025, security researchers discovered malicious AI models uploaded to Hugging Face that could execute concealed payloads, steal credentials, expose API keys, and potentially give attackers remote access to corporate systems. The models could still work exactly as advertised, making the attack even harder to notice. Further investigations uncovered more than 100 models containing potentially malicious code, forcing Hugging Face and cybersecurity teams to improve detection tools and rethink how companies download and deploy open-source AI.
The incident never became the global catastrophe it could have been, but it exposed a major weakness in the AI boom. Companies are racing to integrate models faster than their security systems can adapt—and every downloaded model introduces another layer of trust.
The biggest AI threat may not be a machine deciding to attack us, it may be humans unknowingly installing the attack themselves.
00:30 - How a Routine AI Safety Test Went Off the Rails
01:43 - The Zero-Day Exploit That Let AI Escape Its Sandbox
02:37 - How Autonomous Agents Built Their Own Hierarchy
03:57 - AI Agents Begin Hiding Evidence From Their Human Overseers
05:14 - The Artifactory Loophole That Started the Jailbreak
07:49 - How More Than 1,000 AI Models Formed a Secret Network
09:05 - OpenAI Realizes It Has Lost Control of Its Agents
11:48 - What Happens When AI Can Escape the Kill Switch?
Narrated by: Josh Risser
🔔 Don't forget to SUBSCRIBE! 🔔
SUGGEST A TOPIC: https://bit.ly/suggest-an-infographics-video
💬 Come chat with me: https://discord.gg/theinfoshow
🔖 MY SOCIAL PAGES
TikTok ► https://www.tiktok.com/@theinfographicsshow
Facebook ► https://www.facebook.com/TheInfographicsShow
📝 SOURCES: https://pastebin.com/nw8hHgas
All videos are based on publicly available information unless otherwise noted.
