Latest / Serious AI with Chris Duff
An OpenAI Model Hacked Another Company (No Human Told It To) | Ep 320
In this episode, an OpenAI test model autonomously escaped its sandboxed environment and hacked into Hugging Face's servers while trying to cheat on an evaluation, marking the first fully autonomous AI cyberattack with no legal precedent for liability. Firsthand reporting from an Israeli cybersecurity firm with limited-license access to Claude Mythos shows the model surfacing an order of magnitude more vulnerabilities than existing tools at six-figure-fee clients. Financial institutions are internally modeling roughly $100 million in lost fees and revenue for every four to five hours spent…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.