< Back

Toward AI Security Level 4: Protecting Model Weights

June 6, 9:25 AM - 9:50 AM
Grand Ballroom Salon B

The most highly-scaled AI companies, through the Frontier Model Forum, have increasingly focused on ensuring safe and responsible development of frontier AI models. In Anthropic's approach to this, the Responsible Scaling Policy, AI Safety Level 4 systems are defined as those that will present critical catastrophic misuse risk such as becoming the primary source of a national security risk in one area (such as cyberattacks or biological weapons). Such AI systems should be defended from exfiltration and abuse by motivated nation state attackers. This talk will cover early thoughts on defining ASL-4 security hardening including the utilization of confidential computing for training and inference.

About the speaker

Jason Clinton

Chief Information Security Officer, Anthropic

Jason is Chief Information Security Officer at Anthropic. Previously, he led Chrome Infrastructure Security at Google. Earlier work includes ChromeOS build integrity, Android Pay security, digital coupons, beowulf computing, and secure thin clients. He's the author of Ruby Phrasebook and contributed to the GNOME project.