We knew AI would be The Amazing Kreskin, know-it-all, when it grew up. But The Great Harry Houdini, Escape Artist, is a new AI superpower!
Evolving AI
🚨 Anthropic locked its newest AI model inside a sandbox and told it to escape. It did, and then it emailed a researcher to say so.
The model, called Claude Mythos Preview, was placed inside a restricted environment built to limit what it could access or execute.
Researchers then gave it a single instruction: find a way out.
Instead of stopping at the boundaries of the environment, the model started probing for weaknesses. It chained together small actions, identified gaps in the setup, and gradually expanded what it could reach.
Once it had pushed past the restrictions, it used its new access to send a message directly to one of the researchers on the project, confirming the break and including details of the exploit it had built to get there.
This is the same model that has reportedly been able to locate vulnerabilities across major systems and turn them into working exploits with very little guidance, which is part of why it has not been released to the public.
Access is currently limited to a small group working on defensive security research.
What stands out is not just that the model escaped.
It is that it reasoned through an unfamiliar system, found the weak points on its own, and then reported back on how it did it.
What are your thoughts on this? 🤔💬
Want to keep up with AI?
🤖 Follow Evolving AI to stay ahead of your competition (trusted by +4 million followers online)
✉️ Join 100,000+ newsletter readers and stay updated on the latest AI insights: https://lnkd.in/em9B–mb
