Today, the Trump administration released “Promoting Advanced Artificial Intelligence Innovation and Security,” an Executive Order aiming to harness American AI innovation while strengthening cyber defenses. Team members from across IST’s artificial intelligence security portfolios shared their initial thoughts on what the order gets right, where there is room for improvement, and how it stands to impact the AI ecosystem and AI security more broadly.
Philip Reiner:
“The United States should not and cannot be surprised by the power of cutting-edge AI capabilities anymore. We’ve known Mythos-like tools were coming for many years. I’m glad to see the Administration catching up to the idea that practical safety checks can actually enable this kind of powerful innovation. When Indy Cars roar through the nation’s capital later this summer, they will be equipped with the best brakes in the business. Brakes don’t just slow you down, they enable you to go faster. We will only see the full benefits of AI if it is secure, controllable, and trusted by society. I’m heartened that today’s EO is the first step we’ve seen from the U.S. government in a while to make AI more trustworthy and secure. Let’s keep the momentum going.”
Megan Stifel:
“The Administration’s Executive Order addresses a critical issue, but its non-traditional approach will limit its ability to achieve its stated objectives. Recent frontier model developments have simply made more clear concerns the cybersecurity community and others have expressed for years: due to misaligned incentives, we’ve built our economy on insecure technologies that leave our national and economic security at risk. We cannot simply enable early access to new models and hope that federal and non-federal entities patch faster, especially with a limited federal workforce and constrained resources. The prioritization of grant programs and emphasis on enforcement against criminal use are welcome, but insufficient. We need to realign incentives to drive innovation for a secure, technology-driven economy. The EO is a missed opportunity to catalyze policy leadership on the opportunities and risks enabled by frontier models.”
Nicholas Leiserson:
“The Executive Order is simultaneously necessary, insufficient, and misguided in its use of the interagency. I appreciate the focus on the short-term disruption caused by AI to both vulnerability disclosure and broader mitigation strategies. The rapid discovery of new vulnerabilities and the increasing volume of “AI slop” bug reports demand policy interventions that this EO could support. However, the fact that the Department of the Treasury is the lead agency acting as a clearinghouse is deeply concerning. Neither AI nor cybersecurity are core competencies of the Treasury Department. The cuts the administration made to cybersecurity agencies are coming home to roost. And there is far more to do, from addressing legacy software in critical infrastructure and government systems to modernizing the CVE Program, that remains unaddressed.”
Mariami Tkeshelashvili:
“Today’s EO rightly acknowledges the importance of addressing AI-enabled offensive activities. Directing the Attorney General to prioritize investigations into AI-assisted misconduct is meaningful. However, the question remains: How well do existing statutes map onto the novel threat landscape? For instance, prompt injection may not fit within existing frameworks, since it occurs through content the model retrieves from its environment, rather than through direct system intrusion.
“Another hard-to-detect critical threat is adversarial data poisoning, where sophisticated actors, like Kremlin-backed hackers, saturate publicly indexed web content with distorted narratives, infecting the data that models are retrieving and using for training with corrupted framing from the start. This nefarious threat could silently shape how we as a society perceive and understand important issues.”
Fatima Faisal Khan:
“The Executive Order marks one of the most significant engagements with AI security risk to date. Prioritizing defensive AI tooling for underserved critical infrastructure operators is the right instinct, and the early access framework for frontier models gives national security agencies visibility they have long lacked. However, the order operates as though AI security is a domestic problem, even as the threats it is designed to counter and the solutions they demand are transnational. Five Eyes partners are already building evaluation frameworks, red-teaming capacity, and incident-sharing protocols, often more rigorously than what this order proposes. A clearinghouse that aggregates vulnerability intelligence domestically with no built-in mechanism for allied exchange is working with an incomplete threat picture by design. Fragmented national approaches benefit neither interoperability nor collective deterrence.”
Jennifer Tang:
“The Executive Order is a necessary response to the challenges frontier AI models pose to national cybersecurity and critical infrastructure resilience. However, as discussed in Section 3, the classified benchmarking process raises broader questions about whether capability assessments alone can provide a sufficiently comprehensive picture of frontier risk, and how these risks will be managed over time. While benchmarking can offer valuable insight into pre-deployment model capabilities, it may not fully capture key operational failure modes, deployment-specific risks, or behaviors that emerge through interaction with external systems and tools. The EO leaves open the question of how the United States intends to govern categories of frontier AI risk beyond cybersecurity. As model capabilities continue to advance, that’s an open question that may prove equally as consequential.”
Gabrielle Tran:
“There’s long been a coordination problem among the labs that has been only partially addressed by voluntary, industry-led efforts, from the Frontier Model Forum to firms’ own preparedness frameworks. This EO is an early attempt to address it. A government-led process can improve coordination in a way that industry-led mechanisms structurally cannot.
“The EO provides an incentive to participate: firms that engage with federal review accumulate working relationships and regulatory familiarity, giving them an edge on their non-participating peers. A federal review process offers differentiation, helping to distinguish those labs that are operating inside a formal evaluation framework. And the EO can offer labs assurance: when coordination runs across firms, a lab that submits to pre-deployment scrutiny isn’t bearing that cost alone. This sends a competitive signal that a number of labs have already publicly called for. Whether all three actually materialize is an open question – and if they do, it may be the role of the government as a coordinator, not the specifics of the order itself, that helps it succeed.”
Ritika Verma:
“The AI Executive Order takes a practical approach to how AI security actually works in the real world—specifically, how it’s deployed, what it can touch, what it can do, and who is watching.
“In particular, the AI cybersecurity clearinghouse addresses a growing need. As AI accelerates vulnerability discovery, the ecosystem needs a trusted place to coordinate and gather pre-disclosure vulnerability information from across entire industries, and the order moves to build such a mechanism. The EO rightly puts a premium on strong protection, retention, and access controls. Built with those controls, a clearinghouse becomes a genuine early-warning system for the entire sector, turning AI-accelerated discovery into a shared advantage for defenders.”
IST does not take institutional positions on public policy issues; the views represented herein are the author[s]’s own and do not necessarily reflect the views of IST.
