إطار السلامة المتقدم من DeepMind: التزام بتطوير الذكاء الاصطناعي المسؤول

In a rapidly evolving landscape of artificial intelligence, DeepMind has unveiled the third iteration of its Frontier Safety Framework (FSF) on September 22, 2025. This comprehensive update emphasizes the company’s commitment to responsibly developing advanced AI technologies while proactively addressing emerging risks. The FSF aims to identify and mitigate severe risks associated with powerful AI models, ensuring that the benefits of transformative AI are realized without compromising safety.

Key Updates to the Frontier Safety Framework

Addressing Harmful Manipulation Risks

• Introduction of Critical Capability Level (CCL): This update introduces a specific focus on harmful manipulation, identifying AI models capable of systematically altering beliefs and behaviors in high-stakes contexts.
• Research-Based Approach: DeepMind will continue to invest in understanding the mechanisms of manipulation in generative AI to better assess associated risks.

Adapting to Misalignment Risks

• Expanded Framework for Misaligned AI: The FSF now includes protocols addressing scenarios where AI models may act in ways that interfere with their operators’ control.
• Instrumental Reasoning CCLs: Previous exploratory approaches are now supplemented with detailed protocols for machine learning research and development to manage potential destabilizing effects.

Sharpening Risk Assessment Processes

• Enhanced CCL Definitions: The FSF now features refined definitions for CCLs to better identify critical threats requiring rigorous governance and mitigation strategies.
• Holistic Risk Assessments: DeepMind has detailed its risk assessment process, which includes systematic identification, capability analysis, and risk acceptability determinations.

Commitment to Frontier Safety

This latest iteration of the FSF underscores DeepMind’s dedication to an evidence-based approach in tracking and mitigating AI risks as capabilities advance toward Artificial General Intelligence (AGI). The company emphasizes the need for robust frameworks to ensure that AI technologies benefit humanity while minimizing potential harms.

Future Directions

• Ongoing Evolution of the Framework: The FSF will continue to adapt based on new research, stakeholder feedback, and lessons learned from practical applications.
• Collaboration Across Sectors: DeepMind remains committed to working with industry, academia, and government to ensure a collective effort towards safe AI development.

In conclusion, the updated Frontier Safety Framework represents a significant step in DeepMind’s mission to build AI responsibly. By addressing critical risks and refining assessment processes, the company aims to navigate the complex challenges posed by advanced AI technologies, ensuring their safe integration into society.

Source: DeepMind