A Meta policy document describes the company’s fears that it could accidentally develop an AI model which would lead to “catastrophic outcomes.” It describes its plans to prevent the release of such models, but admits that it may not be able to do so.
Among the capabilities the company most fears are an AI system that could break through the security of even the best-protected corporate or government computer network without human assistance …
TechCrunch spotted the policy document bearing the innocuous-sounding title of Frontier AI Framework.
The document, which Meta is calling its Frontier AI Framework, identifies two types of AI systems the company considers too risky to release: “high risk” and “critical risk” systems.
As Meta defines them, both “high-risk” and “critical-risk” systems are capable of aiding in cybersecurity, chemical, and biological attacks, the difference being that “critical-risk” systems could result in a “catastrophic outcome [that] cannot be mitigated in [a] proposed deployment context.” High-risk systems, by contrast, might make an attack easier to carry out but not as reliably or dependably as a critical risk system.
The company explains its definition of a “catastrophic” outcome:
Catastrophic outcomes are outcomes that would have large scale, devastating, and potentially irreversible harmful impacts on humanity that could plausibly be realized as a direct result of access to [our AI models].
One example given is the “automated end-to-end compromise of a best-practice-protected corporate-scale environment.” In other words, an AI that can break into any computer network without needing any help from humans.
Others are:
- Automated discovery and exploitation of zero-day vulnerabilities
- Fully automated scams against individuals and businesses, causing widespread harm
- The development and proliferation of “high-impact biological weapons.”
The company says that when it identifies a critical risk, it will immediately cease work on the model and seek to ensure that it cannot be released.
Admits containment may not be possible
Meta’s document frankly admits that the best it can do in these circumstances is to do its best to ensure that the model is not released, but its measures may not be sufficient (italics are our emphasis):
Access is strictly limited to a small number of experts, alongside security protections to prevent hacking or exfiltration insofar as is technically feasible and commercially practicable.
You can read the full policy document here.
Photo by Cash Macanaya on Unsplash
FTC: We use income earning auto affiliate links. More.
