John Thickstun

Assistant Professor

John Thickstun, assistant professor of computer science at Cornell University, studies machine learning and has spoken extensively on AI advancement, regulations and infrastructure investment.

Thickstun says: 

“It's hard to see what outcome this EO hopes to achieve. I can see an analogy in this EO to current practices whereby security researchers disclose vulnerabilities to software vendors before making a public announcement, giving them time to fix the vulnerability before it's made widely known. Perhaps the EO is inspired by the recent press around Anthropic's Mythos model, which notably found a bug in OpenBSD. But the proposed EO disclosure to the government isn't directly actionable like a traditional vulnerability disclosure. What's the government proposing to do with a model when it gets access 30 days early? How is it going to act on that early access? What happens if a government employee identifies something concerning about a model?

“Without clearer answers to these questions, my read of this EO is that it creates some appearance of oversight while largely continuing the administration's hands-off approach to AI governance.”

Adrian Sampson

Associate Professor

Adrian Sampson, associate professor of computer science, studies programming languages and computer architecture.

Sampson says:

“The centerpiece of this order is a voluntary 30-day review period. AI labs submit models that might be a cybersecurity risk to the federal government before they ‘release such models to other trusted partners.’ This is a largely meaningless step. It does not address any of the fundamental risks posed by AI-powered vulnerability discovery.

“It might let the government more quickly surface security vulnerabilities in its own infrastructure that are only revealed by a specific new LLM, but the real risks are about the long-term destabilizing effects on all technology infrastructure. In other words, the risks are about what happens over time when the cost of vulnerability discovery goes down and all infrastructure, public and private, becomes more exposed. A 30-day testing window by a federal agency does nothing to address this systemic shift.

“Addressing this kind of systemic, long-term risk requires actual regulation describing how these models can be trained and deployed. A short delay cannot address it.

“I am skeptical of the EO's plan to develop a ‘classified benchmarking process to assess the advanced cyber capabilities of AI models.’ It is fundamentally impossible to understand the cybersecurity capabilities of a newly trained model by evaluating its ability to find already-known vulnerabilities. These are likely to be present in its training set. The proof that one model is more capable than another comes from its ability to find vulnerabilities that no one else has found before. It is hard to see how any ‘benchmark’ could approximate that property.

“It is particularly ironic that this EO comes just days after the same administration announced plans to further dismantle the American scientific system. Part of the EO even requires directing grant money toward ‘advanced AI vulnerability detection.’ While the intersection of AI and cybersecurity is clearly an important area, this provision conveys a model of grantmaking that is reactive and short-term. This is the wrong way to strengthen the security and reliability of the country's technology infrastructure. Doing so requires long-term, sustained investment in the fundamental technical underpinnings of technical resilience. If we as a nation focus our scientific effort on problems that arose in last week's headlines, we will quickly fall behind the evolving threat landscape.”

Ayham Boucher

Lecturer of information science and the executive director of Cornell’s AI Innovation Hub

Ayham Boucher is a lecturer of information science and executive director of AI Strategy and Innovation at Cornell.

Boucher says:

“This EO is a balancing act between two meanings of ‘beating China.’ One meaning is: move fast enough that U.S. companies stay ahead. The other is: use that lead to give the U.S. government and critical infrastructure an early defensive window before adversaries see the same capabilities. This EO tries to do both by creating a voluntary early-access framework, while explicitly avoiding a licensing or preclearance regime.

“The voluntary structure is also a strong trust signal between the U.S. government and the leading AI labs. It shows that the government is betting on the leading AI labs to act like national-security partners without having to force the issue.”