{"type":"video","version":"1.0","html":"<iframe src=\"https://www.loom.com/embed/ac22cd9189944ba592b70a4adbae3896\" frameborder=\"0\" width=\"1280\" height=\"960\" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>","height":960,"width":1280,"provider_name":"Loom","provider_url":"https://www.loom.com","thumbnail_height":960,"thumbnail_width":1280,"thumbnail_url":"https://cdn.loom.com/sessions/thumbnails/ac22cd9189944ba592b70a4adbae3896-5a14bb35882603d2.gif","duration":210.766667,"title":"AI Jailbreaks, Safety Limits, and Resilience","description":"This Loom discusses the safety and technical debate around a US government order to Amphropic to suspend its frontier AI models Fable 5 and Mythos 5. It attributes the directive to a report that a user successfully executed a jailbreak, and compares how jailbreaks and prompt chaining can bypass language-based safety filters. The author explains universal versus non-universal jailbreaks and notes Anthropics argument that perfect jailbreak resistance is physically impossible, while Amphropics defense focuses on making universal breaks harder and enforcing strict 30-day user data retention to monitor attacks. The Loom concludes with concerns about “zero tolerance” decisions versus the industry’s risk-management approach and emphasizes building trustworthy, resilient AI."}