Necessary Measures and Policy Focus on AI Governance

Significant changes in AI governance are happening at the federal level, but the discussion is ongoing. In June, the White House issued an executive order focusing on AI innovation and security, along with National Security Presidential Memorandum-11 concerning AI in national security. This happened as new AI models acquired advanced cyber capabilities.

Before its review framework was functional, the administration restricted Anthropic’s Fable 5 and its Mythos model under export-control authority. OpenAI, similarly, withheld the release of GPT-5.6 pending government approval. Meanwhile, agencies and companies are rapidly integrating AI agents into daily operations.

Delegation and Trust in AI

Agentic AI concerns delegation. AI systems capable of drafting emails, searching databases, filing forms, writing code, monitoring networks, or routing requests are now trusted to act, often covering several steps before human oversight. These systems are improving at a rapid pace. The Model Evaluation and Threat Research organization monitors their progress, comparing it to tasks done by human experts. In 2025, AI performance was doubling every seven months; now, it seems to be doubling every four months.

Building Governance and Capacity

Organizations must build the ability to govern AI agents while human control remains feasible. Proper deployment of agents can transform citizen-government relations. For instance, a small business could focus more on customers and less on bureaucracy, while veterans could speed up benefit claims processing. Agencies might use agents to streamline processes, reduce backlogs, and improve service quality.

However, these benefits depend on trust, reliability, and security. Poorly governed agents risk sharing information incorrectly, exceeding authority, or hiding errors in automated chains. This could result in misallocated benefits, damage to infrastructure, or conflict escalation. AI decision-support systems are currently used for military targeting recommendations, but guidance and technical standards are lacking.

Challenges in Trustworthy Delegation

Much AI policy debate centers on access—who accesses models, chips, data, and energy. As AI systems act, decisions on trustworthy delegation arise. Policymakers need reliable, secure, and accountable use methods, understood by both government and public.

Establishing responsible use means having trained staff, sound procurement, clear authority lines, audit logs, and decision reconstruction capabilities. Deployment infrastructure will determine if AI strengthens public institutions or weakens them.

Focus on Cybersecurity

The Mythos model by Anthropic, excelling in identifying software vulnerabilities, shows agentic capabilities’ rapid development for both defense and attack. Industry initiatives like Anthropic’s Project Glasswing and OpenAI’s Daybreak involve giving vetted defenders access to advanced tools, through differential access.

The White House prioritizes access, but this won’t aid hospitals, utilities, and agencies exposed to cyber threats, unless they possess the necessary staff, standards, and practices. The Cybersecurity and Infrastructure Security Agency and the National Security Agency, with allies, recently released guidance on agentic AI services. Priorities include investing in evaluation and audit capacity, and understanding multi-agent interactions under stress.

Policy and Export Controls

The June executive order tested requesting developers to present capable models for review before release. Evaluation science needs to keep up with evolving models. The Center for AI Standards and Innovation oversees much of this work, collaborating with entities like OpenAI and Anthropic. Despite this, its $10 million budget limits its potential. Congressional support could enhance its evaluation expertise, complementing the NSA’s abilities in national security deployments.

Policymakers also ought to reinforce export controls to secure U.S. advantages in AI. The emphasis isn’t only on having capable models, but shaping agentic systems with aligned values. Recent laws like the Chip Security and Stop Stealing Our Chips Acts aim to maintain the U.S.’s lead.

AI agents remain controllable today, a starting point as current deployments establish long-term institutional habits. Future AI policy should not only focus on powerful AI access but on creating governance frameworks for reliable, accountable, and secure usage.

Jenny Marron, Executive Director of the Institute for AI Policy and Strategy, previously worked at the White House National Security Council and the U.S. Department of State.

Gauthier