AI safety researcher Yo Shavit warns OpenAI researchers on visas face losing internal model access within two months

@TheZvi @allTheYud Even green card holders and US persons are restricted from participation if the project they are working on has a high enough national security level. They are barred from working on it per Executive Order 12968 signed by Bill Clinton in response to Aldrich Ames.

@yonashav Why should they expect that? This measure won't be used against OpenAI.

@TheZvi *many employees to lose access (most labs have an available pool of US-person research leadership that could e.g. keep an RSI-loop going)

@yonashav OpenAI and its proxies have paid off the administration to the tune of tens of millions of dollars, and have gone out of their way to pay lip service to the admin and show a willingness to do whatever the admin asks.

@yonashav You're assuming this wasnt just punitive politics at play from trump admin

@AndrewCurran_ @TheZvi @allTheYud They almost certainly haven’t made it classified work

It’s near impossible to legally do that with public domain work not derived from other classified information fwiw

Probably just ear listing

@TheZvi We are entering the age of major, permanent gap between models that are available only internally (used for RSI and maybe some applied products w/o API), vs. published heavily hobbled models.

Much harder to regulate internal models.

@TheZvi

@yonashav

@TheZvi I don’t know the answer to whether those authorities exist; my understanding from @TimSchnabel ‘s tweets is that the authority required to target model outputs cannot be used for US persons.

(They could use non-EC authorities, but those probably aren’t model-specific.)

You are referencing a security clearance requirement.

The Anthropic order came from the Commerce Department which has nothing to do with this. The order is likely subject to the EAR and under that “foreign person” / “foreign national” does not include U.S. citizens or green-card holders.

@yonashav Fable was only banned because Anthropic has been screaming about how their models are too dangerous to exist and have the ability to fell nation-states.

Has OpenAI said anything of the sort recently?

No. Stop fear mongering.

@yonashav Move them to London offices.

@yonashav geopolitically, who will be the winner here?

@AndrewCurran_ @TheZvi @allTheYud Ok but the manhattan project wasn’t limited to American citizens. If the us does nationalize a lab it will figure out how to grandfather all researchers it needs - they just will likely be travel restricted.

@kristoph @TheZvi @allTheYud I'm talking about post-nationalization, when capabilities work becomes a government project requiring clearance.

@Tylerkaerr @yonashav Actually, the legal arguments may well result in a broader directive that affects frontier models in general which would include OAI. If Ant’s team argues this is targeting and Gpt 5.5 is susceptible to the same jailbreak, it’s not far fetched to see the revision

@bvhughes @yonashav Anthropic is not the arbiter of truth on all models, only their own. OpenAI is the more definitive authority regarding their own models. OpenAI will simply say that their models are not dangerous. Whatever Anth might say about OAI’s models is almost irrelevant.

@Tylerkaerr @yonashav It’s not about them being anything, it’s going to more than likely come down to legal arguments just like Amazon saying something about the jailbreak doesn’t make them the authority on Ant’s models.

@bvhughes @yonashav The legal argument is pretty simple. The government will take the company creating and releasing any model as authoritative on the subject of their own model when it comes to safety.