Those that argue for more government oversight on AI are typically signaling that they are increasingly falling behind in the AI race. Now granted, we do need some form of government assessment process. I’d however be incredibly weary of any type of a Meta style elder statesmen overview commission. Full of retired non-technical politicians who are going to be easy to influence.

The best way to structure this is to have the intelligence community set up a red team and do a second opinion on new models. If you think about the type of person both in terms of technical capability and interest to do this work thoroughly, then the most logical place in the US government is the intelligence community. Another logical place are the national labs that typically employ highly technical folks. Again, there will be interest there because AI deeply touches work there. This is both salient for closed and open source models. Salient for models hosted in and outside the US.
The intelligence community and/or the national labs should provide a second opinion on the red team assessment and iterate with commercial labs to resolve any security issues.

In terms of the realistic threats of these models, again nuance is important. These models aren’t going to allow teenagers in Russia or Iran with decent coding skills to become super hackers. The risk is that they make already capable actors more effective. Those actors are likely to find social engineering ways to get access to closed sourced models. The most potent tool here is statistical usage monitoring. Flagging suspicious use that falls outside typical patterns. The link here with the intelligence community is salient as they’d probably be interested in these cases and would want a say in how to deal with them. The lets just set up some commission for this misses the important nuanced of how this should be dealt with, both from a competitiveness and security standpoint.