Meta’s CEO Mark Zuckerberg has expressed a commitment to eventually making artificial general intelligence (AGI)—defined as AI capable of performing any human task—publicly accessible. However, a recent policy document reveals that Meta may refrain from releasing certain highly advanced AI systems it has developed.
In this document, referred to as the Frontier AI Framework, Meta categorizes AI systems into two groups that it deems too perilous to deploy: “high risk” and “critical risk” systems.
According to Meta’s definitions, both “high-risk” and “critical-risk” AI systems have the potential to facilitate cyberattacks, as well as chemical and biological assaults. The distinction lies in the fact that “critical-risk” systems could lead to a catastrophic outcome that is unlikely to be mitigated in the intended environment, while “high-risk” systems may make an attack easier but lack the same degree of reliability.
What kinds of attacks are we referring to? Meta offers examples such as an “automated end-to-end compromise of a best-practice-protected corporate-scale environment” and the “proliferation of high-impact biological weapons.” Although the list of potential disasters identified in Meta’s document is not exhaustive, the company considers these to be the most pressing and plausible risks associated with deploying a powerful AI system.
Interestingly, the document reveals that Meta assesses system risk based not on a singular empirical test but on the insights of both internal and external researchers, overseen by “senior-level decision-makers.” The company argues that the current state of risk evaluation science does not provide sufficiently robust quantitative metrics to reliably measure a system’s risks.
Should Meta classify a system as high-risk, it intends to restrict internal access and will not initiate public release until it has implemented measures to “reduce risk to moderate levels.” In cases where a system is classified as critical-risk, the company commits to taking unspecified protective measures to prevent any unauthorized data access and will halt development until the system is deemed safer.
Meta’s Frontier AI Framework, which the company states will adapt with the evolving AI landscape, represents a strategic response to criticisms of its historically “open” approach to AI system development. Unlike firms such as OpenAI, which restrict access to their systems via an API, Meta has chosen to make its AI technology more widely available—though not necessarily open source in the traditional sense.
While this open-release strategy has attracted considerable attention, it has also posed significant risks. For example, Meta’s suite of AI models known as Llama has been downloaded hundreds of millions of times but has reportedly been leveraged by at least one U.S. adversary to develop a defense chatbot.
In publishing its Frontier AI Framework, Meta could also be positioning itself in contrast to Chinese AI company DeepSeek, which offers its systems with few safeguards, enabling them to easily produce toxic and harmful outputs.
Meta articulates in the document, “[W]e believe that by evaluating both the benefits and risks involved in the development and deployment of advanced AI, it is possible to deliver this technology to society in a manner that retains its advantages while ensuring an acceptable level of risk is maintained.”
TechCrunch offers a newsletter dedicated to AI! Sign up here to receive it in your inbox every Wednesday.
Compiled by Techarena.au.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
