Anthropic Loosens AI Safety Policy

🔑 Enhanced Key Takeaways

• •Anthropic's original 2023 RSP categorically barred training AI models above certain capability levels without pre-existing adequate safety measures, a restriction now removed in version 3.0[1].
• •The updated RSP commits to greater transparency via public Risk Reports on model safety testing and requires external expert reviews for high-risk assessments[3][5].
• •Anthropic will delay AI development only if it leads the field and perceives significant catastrophe risks, while pledging to match or exceed competitors' safety efforts[1].

🛠️ Technical Deep Dive

• •ASL-3 protections include safeguards like Constitutional Classifiers, access controls for trusted users, red-teaming, bug bounties, and threat intelligence to counter jailbreaks[2].
• •ASL-3 deployment standards focus on blocking chemical, biological, radiological, and nuclear (CBRN) risks using sophisticated input/output classifiers[3].
• •New capability thresholds added in 2025 include AI R&D-4 (full automation of entry-level AI research) and thresholds for CBRN uplift in state programs[4].

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline