Analysis: AI Models Can Re-derive Cybersecurity Exploits, but May Falter on Specialized Hardware

Many AI models are described as capable of re-deriving cybersecurity exploits. The analysis also considers whether models refuse requests related to dangerous activities, noting that access to controlled materials and physical infrastructure can remain a separate constraint.

The assessment says AI is used mainly to find vendors and suppliers and that its user rarely encounters refusals because they use AI infrequently. AI may be less useful for highly specialized work, such as building a particle accelerator, and it remains unclear how far a model could get with such a task.

The assessment concludes that this capability does not represent a significant proliferation concern for nuclear technologies.