OpenAI

v2 (April 2025) restructured tracked categories to Bio/Chem, Cyber, and AI Self-Improvement, retiring persuasion as a standalone category. Replaced tier system with capability-threshold definitions per category.

• ·Restructured to three categories: Bio/Chem, Cyber, AI Self-Improvement
• ·Retired persuasion as a tracked category
• ·Replaced four-tier system with per-category capability thresholds
• ·Added Research-Stage vs Deployed-Stage thresholds

Original Preparedness Framework beta. Established the four tracked categories (cybersecurity, CBRN, persuasion, model autonomy) and pre-mitigation / post-mitigation tier definitions.

• ·First public Preparedness Framework (beta)
• ·Defined Low / Medium / High / Critical tiers per category
• ·Required Safety Advisory Group review before deployment of high-risk models