Uploaded August 2026 | Updated September 2026, 2 weeks ago

---

At Black Hat, OpenAI gave a talk with preliminary results of their investigation into the Hugging Face incident. And besides multiple zero-days exploited by agents, the agents communicated using a message board. It is absolutely insane and probably the most scary thing I have heard about AI so far.


LEARN ON HEXTREE (ad)
Learn hacking on Hextree: hextree.io
Learn more about AI on Hextree: app.hextree.io/map/artificial-intelligence
Join the Hextree Discord: discord.gg/xgQpCQCpvy

Relevant links:
- First video: youtube.com/watch?v=q2KCrmQz9WE
- hacktron.ai/blog/here-is-how-openai-model-hacked-huggingface
- huggingface.co/blog/security-incident-july-2026
- openai.com/index/hugging-face-model-evaluation-security-incident
- github.com/sunblaze-ucb/exploitgym
- docs.jfrog.com/releases/docs/artifactory-self-managed-releases
- github.com/huggingface/dataset-viewer/pull/3367

CHAPTERS
00:00 - Intro and Recap
03:28 - About the OpenAI Sandbox
07:20 - Agents started a Message Board
11:51 - Reinforcement Learning
16:10 - This is scary!

SUPPORT
Per video: patreon.com/join/liveoverflow
Per month: youtube.com/channel/UClcE-kVhqyiHCcjYwcpfj9w/join
Buy my handwriting font (ad): shop.liveoverflow.com

WATCH, FOLLOW & READ
Second channel: youtube.com/LiveUnderflow
Twitch: twitch.tv/LiveOverflow
Twitter: twitter.com/LiveOverflow
Instagram: instagram.com/LiveOverflow
TikTok: tiktok.com/@liveoverflow_
LiveOverflow blog: liveoverflow.com
Hextree blog (ad): hextree.io/blog

#ArtificialIntelligence #AISecurity #LiveOverflow

(ad) LiveOverflow YouTube channel is supported by advertisement and product placement.