About Executive summary Managing risks Issues Stakeholders Transparency and remedy Annex We also published our Frontier AI Framework, focused on the most critical risks in the areas of cybersecurity threats and risks from chemical and biological weapons. Our framework outlined a number of processes we follow to anticipate and mitigate risk when developing frontier AI systems, including: Identifying catastrophic outcomes to prevent Our framework identified potential catastrophic outcomes related to cyber, chemical and biological risks that we strive to prevent. It focused on evaluating whether these catastrophic outcomes were enabled by technological advances and, if so, identifying ways to mitigate those risks. Threat modeling exercises We conducted threat modeling exercises to anticipate how different actors might seek to misuse frontier AI to produce those catastrophic outcomes, working with external experts as necessary. These exercises were fundamental to our outcomes-led approach. Establishing risk thresholds We defined risk thresholds based on the extent to which our models could facilitate the threat scenarios. We have processes in place to keep risks within acceptable levels, including applying mitigations. We also continued to support a broad ecosystem for responsible AI development and deployment. To empower developers to build safe applications, we expanded our investment in system-level safeguards, including Llama Guard 4 for unified safeguards across text and image modalities and Llama Prompt Guard 2 for improved jailbreak and prompt injection detection. Our work was informed by wide-ranging engagements and consultations with civil society, international organizations, academia and the public on the human rights dimensions of AI. This included Community Forums — conducted in partnership with Stanford University's Deliberative Democracy Lab — that gathered diverse perspectives on the future of AI agents and their governance. 2025 Frontier AI framework Anticipate Evaluate and mitigate Decide Identify reference class Conduct AI risk assessment Assess residual risk Run threat modeling exercises Evaluate for performance and safety Make decision on release Implement mitigations 2025 Meta Human Rights Report 20