Mark Zuckerberg, the CEO of Meta, has promised that one day artificial general intelligence (AGI), which is loosely defined as AI that can perform every work that a person can, will be publicly available. However, Meta notes in a new policy statement that it may not distribute a highly proficient AI system that it built internally in some circumstances.
The company’s concerns that it may inadvertently create an AI model that would result in “catastrophic outcomes” are detailed in a Meta policy paper. Although it acknowledges that it might not be able to stop the publication of such models, it outlines its efforts to do so.
An AI system that could breach the security of even the most secure business or government computer network without human intervention is one of the capabilities that the corporation is most concerned about.
The paper, which Meta is referring to as its Frontier AI Framework, lists two categories of AI systems that it deems too dangerous to make public: “high risk” and “critical risk” systems.
According to Meta, both “high-risk” and “critical-risk” systems may support chemical, biological, and cybersecurity assaults; the distinction is that “critical-risk” systems have the potential to produce a “catastrophic outcome [that] cannot be mitigated in [a] proposed deployment context.” Comparatively speaking, high-risk systems may make an assault simpler to execute, but they are not as dependable or consistent as critical risk systems.
The business defines a “catastrophic” consequence as follows:
Catastrophic outcomes are those that may realistically occur as a direct result of access to [our AI models] and would have significant, destructive, and perhaps irreversible negative effects on mankind.
What kind of assaults are we discussing here? Examples provided by Meta include the “proliferation of high-impact biological weapons” and the “automated end-to-end compromise of a best-practice-protected corporate-scale environment.” Although the business admits that the list of potential disasters in Meta’s document is far from all-inclusive, it does contain those that Meta considers to be “the most urgent” and likely to occur as a direct result of deploying a potent AI system.
Surprisingly, the paper states that Meta categorizes system risk based on the opinions of both internal and external academics, who are then reviewed by “senior-level decision-makers,” rather than on any one empirical test. Why? According to Meta, the science of assessment is not “sufficiently robust as to provide definitive quantitative metrics” for determining the riskiness of a system.
According to Meta, if a system is deemed high-risk, internal access will be restricted, and the system won’t be made public until mitigations have been put in place to “reduce risk to moderate levels.” However, Meta claims that if a system is judged to be critical-risk, it will suspend development until it can be rendered less risky and will put in place unidentified security measures to stop the system from being exfiltrated.
It seems that Meta’s Frontier AI Framework, which the business claims will change with the AI landscape and that it had previously promised to disclose before this month’s France AI Action Summit, is a reaction to criticism of the firm’s “open” approach to system development. In contrast to businesses like OpenAI that choose to gate their systems behind an API, Meta has adopted a policy of making its AI technology publicly available, even if it is not open source by the widely accepted definition.
The open release strategy has been both a boon and a bane for Meta. The company’s Llama family of AI models has been downloaded hundreds of millions of times. However, at least one U.S. enemy has apparently employed Llama to create a defensive chatbot.
It’s possible that Meta wants to compare its open AI approach to that of Chinese AI company DeepSeek by releasing its Frontier AI Framework. Additionally, DeepSeek makes its technologies publicly accessible. However, there aren’t many controls in place, and the company’s AI is readily manipulated to produce hazardous and poisonous results.
The document states that Meta believes “it is possible to deliver that technology to society in a way that preserves the benefits of that technology while also maintaining an appropriate level of risk” by taking into account both risks and benefits when deciding how to develop and implement advanced AI.
The “automated end-to-end compromise of a best-practice-protected corporate-scale environment” is one example cited. To put it another way, an AI that can infiltrate any computer network without human assistance.
Others are:
- Automated zero-day vulnerability identification and exploitation
- Completely automated frauds against people and companies that cause severe damage
- creating and distributing “high-impact biological weapons.”
According to the corporation, if it finds a serious danger, it will stop working on the model right away and try to prevent its distribution.
It might not be possible to limit admissions.
Though its efforts may not be enough (italics are our emphasis), Meta’s whitepaper candidly acknowledges that the most it can do in these situations is to try its best to prevent the model from being released:
In addition to security measures to stop hacking or exfiltration to the extent that they are technically and financially possible, access is rigorously restricted to a small group of specialists.