AI Agents Uncover Dangerous Zero-Day Vulnerabilities

Transcript

1. 0:00We assembled teams of agents powered by our most cyber capable models, tasked them with finding weaknesses in our research environments. Um
2. 0:10While they found no route to the internet they could easily they could easily exploit while concealing their
3. 0:16activity from our monitors, they did find zero-day vulnerabilities. So, zero-day vulnerabilities are like, you need to fix it now.
4. 0:24Those kind of vulnerabilities, right? Like 10 out of 10 risk kind of vulnerabilities. They did find zero-day vulnerabilities,
5. 0:32virtual machine escapes, hardware side channels for communications between agents. Um we used those discoveries to further
6. 0:41strengthen our defenses and contribute fixes to open-source software. Smart. After strengthening our containment
7. 0:48architecture, we resumed our frontier um reinforcement learning training and inference on October 1st. >> Watch the full replay at
8. 0:56community.thesalon.ai.