3 min Security

Meta will not disclose high-risk and highly critical AI models

Meta will not disclose high-risk and highly critical AI models

Meta has established new rules under which it will never disclose internally developed, potentially (very) high-risk AI models. These are two categories of AI models that could potentially cause a lot of harm and possibly be uncontrollable.

Although Meta has always been more or less “open” from the beginning of its AI development work, making its AI technology available to everyone, the tech giant now wants to establish guidelines, especially when it comes to internally developed AI models that may be high-risk.

In its recently published Frontier AI Framework document, which lays out these guidelines, Meta states that these internally developed AI models with high capacity and high to extremely high risk will never be made public.

Two categories

Meta has defined two categories for these high- to extremely high-risk models or systems. The first category, high-risk, involves high-value AI models or systems that can help develop and execute cyberattacks and chemical or biological attacks.

The second category, critical risk, includes the high-risk models and systems that can lead to attacks with “catastrophic outcomes” that cannot be countered when deployed within a given context.

The difference with the first category is that high-risk models or systems make an attack easier to execute but are not as reliable or dependent as a critical-risk model or system.

Examples of these attacks include large-scale, automated end-to-end cyber attacks on companies that have the best protection measures in place or the proliferation of biological weapons.

Human assessment

Meta also explains how it places its AI systems in these categories in the document. In doing so, the systems are not subjected to empirical testing but assessed by a team of internal and external researchers appointed by the social media and tech giant’s management.

Experts determine which AI models and systems contain high risks, according to Meta, because the company believes that evaluating AI models for risk assessment is not yet robust enough to yield reliable metrics.

When Meta determines that an AI model is high-risk, it limits internal access and does not release it publicly until constraints are added that lower the risk to a more “moderate level.

However, when a system is deemed a critical risk, unspecified measures are taken to ensure that the model or system does not become public, and research will be halted until it is made less dangerous.

The Meta Frontier AI Framework will continue to be expanded and modified as AI technology advances.

Read more: Meta wants to flood social media with AI

Response to DeepSeek

Experts say the social media and tech giant’s Frontier AI Framework is a response to the criticism it is receiving for its “open-source” strategy for AI models, writes TechCrunch.

Among other things, Meta is making its Llama models somewhat open source. This allows many developers to use these models for their applications, but not always correctly. For example, a U.S. geopolitical adversary allegedly used the Llama models to develop a military chatbot.

Moreover, creating a document with guidelines for AI models and systems could be a response to the recent developments surrounding China’s DeepSeek. This AI developer makes all its models truly open source, but has few security guidelines, and using this model could lead to harmful outcomes.

Also read: DeepSeek is unsafe, and that has nothing to do with China