Mark Zuckerberg, CEO of Meta, has expressed his commitment to making artificial general intelligence (AGI) openly available in the future. AGI refers to an advanced AI system capable of performing tasks on par with human intelligence. However, a newly released policy document from Meta outlines situations where the company may withhold access to its most powerful AI models due to potential risks.
This policy, known as the Frontier AI Framework, introduces two categories of AI systems that Meta deems too dangerous for unrestricted release: “high-risk” and “critical-risk” AI models.
Defining High-Risk vs. Critical-Risk AI
Meta classifies high-risk AI as systems that could facilitate cybersecurity breaches or contribute to chemical and biological threats, though not in a completely reliable manner. In contrast, critical-risk AI is described as technology that could lead to catastrophic consequences with no effective mitigation in place.
The company provides specific examples, such as an AI capable of executing a full-scale cyberattack on a secure corporate environment or one that could accelerate the creation of high-impact biological weapons. While these are not the only possible threats, Meta identifies them as among the most urgent and realistic dangers posed by advanced AI.
Meta’s Approach to Risk Evaluation
Rather than relying on a standardized empirical test, Meta assesses risk based on expert opinions from both internal and external researchers, with final decisions made by senior leadership. The company argues that current scientific methods for evaluating AI safety are not yet reliable enough to produce definitive, quantifiable risk assessments.
If an AI system is classified as high-risk, Meta will restrict internal access and delay its release until appropriate safeguards are in place. If a system falls under the critical-risk category, the company will implement strict security measures to prevent unauthorized access and halt further development until risks are sufficiently reduced.
Balancing Open AI Development with Responsible Deployment
The Frontier AI Framework is designed to adapt to evolving AI risks and aligns with Meta’s recent pledge to release it publicly ahead of the France AI Action Summit. This move is likely a response to concerns regarding Meta’s historically open approach to AI development. Unlike competitors such as OpenAI, which control access to their models via APIs, Meta has positioned itself as a leader in open AI research, although its models are not fully open-source by traditional definitions.
Meta’s Llama family of AI models has been widely downloaded and used by millions. However, this openness has had drawbacks, including reports that one U.S. adversary leveraged Llama to build a defense-related chatbot.
By outlining clear restrictions in its Frontier AI Framework, Meta may also be distinguishing itself from companies like China’s DeepSeek, which offers freely available AI models with minimal safeguards. Meta emphasizes that balancing innovation with responsible risk management is key to ensuring AI benefits society while maintaining safety.
The document states:
“By evaluating both the advantages and potential dangers of AI deployment, we aim to provide cutting-edge technology in a way that maximizes benefits while keeping risks at an acceptable level.”
This policy underscores Meta’s evolving stance on AI safety—one that seeks to maintain openness while taking measured precautions against potentially harmful applications of its technology.
This version preserves the key ideas but rephrases them to ensure originality while keeping the information clear and engaging.
