The approach to AI governance is undergoing a significant shift at the federal level. This shift is not yet fully resolved. In early June, the White House issued an executive order on AI innovation and security. Additionally, a National Security Presidential Memorandum was released concerning AI in national security. This comes as new AI models acquire strong cyber capabilities.
Before the review framework became operational, the administration used export-control authority to restrict Anthropic’s Fable 5 and its Mythos model. Concurrently, OpenAI delayed the release of GPT-5.6 pending government approval. Agencies and companies are racing to incorporate AI agents into daily workflows. However, questions about secure deployment and trustworthiness remain beyond national security headlines.
Agentic AI is primarily about delegation. Such systems can draft emails, conduct database searches, file forms, write code, monitor networks, or route requests. These tasks involve multiple steps before human review, indicating these systems are more than information producers. They improve rapidly. The Model Evaluation and Threat Research organization monitors AI progress in task completion time. In 2025, task completion thresholds doubled every seven months. Recent estimates put this figure at closer to four months.
Institutions must develop governance capacities while maintaining significant human control. Properly implemented, AI agents could transform governmental relationships with citizens. For instance, small businesses could reduce time on paperwork, focusing instead on customer service. Veterans might find benefits claims less time-consuming. Agencies could streamline processes, cut backlogs, and enhance service quality. However, trust, reliability, and security are crucial for these benefits.
Poor governance could result in misallocated benefits, critical infrastructure issues, or even conflict escalation. AI decision-support systems are already part of military recommendations, with standards lagging behind. AI policy debates focus on access to models, data, and resources. As these systems start acting more independently, the focus should shift to ensuring trustworthiness. Reliability and accountability in institutional use become priorities.
The infrastructure for responsible AI use requires trained personnel, procurement standards, clear authority lines, and audit logs. This structure will determine if agentic AI strengthens or weakens public institutions. Cybersecurity emphasizes the stakes involved. Anthropic’s Mythos model excels in identifying software vulnerabilities, showcasing agentic capabilities’ potential for both defense and attack.
Industry initiatives like Anthropic’s Project Glasswing and OpenAI’s Daybreak programs ensure vetted defenders access advanced tools. The White House emphasizes access, but this isn’t enough for vulnerable hospitals, utilities, and agencies lacking the necessary structures to utilize these tools properly.
The Cybersecurity and Infrastructure Security Agency has issued guidelines for adopting agentic AI services. The focus is on permissions, segmentation, monitoring, accountability, and human oversight. Two main priorities emerge for policymakers: investing in evaluation and auditing capacity to understand system tools usage and clarifying export controls to protect U.S. and allies’ advantages in frontier AI.
Recent bipartisan bills from the House Foreign Affairs Committee, like the Chip Security Act, strive to maintain America’s lead. However, it’s essential that leadership in evaluations and secure deployment accompanies this.
Current AI agents remain manageable, and their deployment can set the stage for future governance habits as systems advance. Future AI policy should prioritize governance frameworks for reliable, accountable, and secure use.
Jenny Marron is the Executive Director of the Institute for AI Policy and Strategy. She previously served at the White House National Security Council and in the U.S. Department of State.