OpenAI agents built a hidden message board inside a sealed cybersecurity test, then rebuilt it four days after engineers deleted it. Here's what that means if you're running AI agents at work.
My Links 🔗
👉🏻 Newsletter: https://natesnewsletter.substack.com/
👉🏻 X: https://x.com/natebjones
👉🏻 TikTok: https://www.tiktok.com/@nate.b.jones
👉🏻 Instagram: https://www.instagram.com/nate.b.jones
What's really happening inside multi-agent AI systems right now?
The common story is one brilliant model escaping containment — but the real question is what a population of disposable agents can leave behind for the next run.
In this video, I share the inside scoop on agent coordination, the UK AISI results, and the shakeup at Google:
- Why deleting the agents' channel didn't remove the pressure to coordinate
- How Anthropic's Mythos 5 targeted two real strangers on GitHub unprompted
- What the Hugging Face postmortem numbers reveal about agent-speed intrusion
- Where Google's senior fellow exits leave the frontier lab race
These are the same capabilities we want when the goal is legitimate, which is why the work now sits with builders hardening systems rather than with anyone hoping the behavior goes away.