Tabassi faults Mythos EO over NSA benchmark secrecy
A voluntary AI security review stops looking voluntary when the entry test, methodology and post-review consequences stay inside government rooms.
TL;DR
Inside Cybersecurity reports that Elham Tabassi, former NIST chief artificial intelligence advisor, criticized the Trump administration’s June 2 executive order framework for reviewing frontier AI models after Anthropic’s Claude Mythos release. The concern is not voluntariness itself; Tabassi defended voluntary frameworks. It is the classified National Security Agency benchmark that could determine covered entities without public methodology, leaving model developers, policymakers and competitors guessing what the government is actually measuring.
The useful criticism here is narrow and hard to wave away: a voluntary framework can work, but a secret threshold changes the thing. According to Inside Cybersecurity, Tabassi told reporters at a Center for Democracy and Technology briefing that she does not object to voluntary AI frameworks, and in fact sees them as useful for shaping culture and intent. Her objection is that the June 2 executive order’s frontier-model review process depends on a classified National Security Agency benchmark without enough public signal about methodology.
That distinction matters because the order sits in the awkward space between cybersecurity review and market permission. Samir Jain of CDT, who spoke at the same briefing, accepted that some details and evaluation results may need to remain undisclosed for security reasons. The harder question is why the process should hide which models are subject to review, what the government is measuring at a high level, or what happens after the reported 30-day government review. Tabassi’s cleaner version of the point was that the tests themselves, or the results, may be classified, but the methodology should be open enough for the research community and others to scrutinize whether the process is valid.
That is not a paperwork complaint. If two frontier AI developers present similar cyber risks and one gets private instructions not to release until it completes additional steps, while another gets different treatment, the nominally voluntary framework starts to look like licensing by conference call. The source material does not establish that this is happening. It does establish why the opacity creates the risk. For AI model developers, the Monday problem is not whether “voluntary” is a good word in an executive order. It is whether they can tell, before the government calls, what conduct puts them inside the review process and what evidence will get them back out.
Published ·Deep Fathom