A new presidential memorandum establishes a mechanism that could remove AI companies prioritizing safety guardrails from U.S. defense and intelligence contracts. National Security Presidential Memorandum 11 (NSPM-11), released in early June 2026, includes language allowing federal agencies to terminate contracts with companies demonstrating "a pattern of conduct inconsistent" with the government's AI policies, though renewable one-year waivers could indefinitely delay enforcement.

What Does NSPM-11 Actually Require?

The memorandum establishes four pillars for AI adoption in national security: adoption, adaptation, assurance, and accountability. The document's most consequential provision states that "no commercial entity or adversary possesses the capability to prevent use of, disable or degrade, or materially modify without Federal Government knowledge and approval, an AI system that our men and women depend on for their missions." According to analyst Zvi Mowshowitz, this language means once a company delivers an AI system to the government, that company cannot enforce its own safety policies or contractual restrictions on how the system is used.

The memorandum also extends restrictions to subcontractors, preventing defense contractors from using technology from companies that maintain safety restrictions as part of larger systems. Section 3 directs the Department of Defense to update its AI directive annually and authorizes agency heads to terminate contracts with non-compliant vendors.

How Could This Target Safety-Focused Companies?

According to Mowshowitz's analysis, the memorandum's language and timing suggest it could be used to exclude Anthropic, a company known for refusing to remove safety features from its AI models even when government officials request it. The document does not name Anthropic explicitly, but Mowshowitz notes that the provision allowing waivers "seem reasonably likely to happen" indefinitely, potentially making enforcement more theoretical than practical.

Anthropic has built its business model around the principle that AI safety is non-negotiable. The company has repeatedly resisted demands to disable safety guardrails, arguing that unrestricted AI systems pose genuine risks. This principled stance creates friction with defense officials seeking maximum flexibility in AI deployment.

"This seems like a solidly smart policy document," remarked Dean W. Ball in response to the memorandum, though Mowshowitz's interpretation emphasizes the potential for excluding safety-focused vendors.

Dean W. Ball, Policy Commentator

What Are the Competing Visions for AI Governance?

The memorandum reveals a fundamental disagreement about managing advanced AI systems. Anthropic and researchers at Google DeepMind have called for international coordination to potentially slow down AI development in the name of safety. The U.S. government, by contrast, signals it wants to accelerate AI adoption across military and intelligence operations without restrictions.

OpenAI has proposed a different approach, calling for international coordination among key actors to ensure the ability to slow down AI development while also warning about the dangers of concentrated AI power. However, OpenAI's own plans explicitly include recursive self-improvement, where AI systems improve themselves without direct human oversight.

"OpenAI's plan is, very explicitly, AI to go into recursive self-improvement. I appreciate the honesty, but the inherent contradictions remain, and are not addressed," noted Zvi Mowshowitz.

Zvi Mowshowitz, AI Policy Analyst

How to Understand the Practical Implications of This Policy

• Market Incentive Shift: The memorandum signals that AI companies prioritizing safety over government access will lose significant revenue opportunities in the defense sector, potentially shifting industry incentives toward less cautious approaches to AI development.
• Waiver Mechanism Uncertainty: While the memorandum allows one-year renewable waivers for "operational imperatives, test and evaluation arrangements, threat intelligence sharing, and other mission-critical applications," Mowshowitz suggests these waivers appear likely to be used indefinitely, making the restriction potentially theoretical in practice.
• International Coordination Tensions: The memorandum suggests the U.S. government is moving away from international agreements to coordinate AI development, instead prioritizing unilateral capability advantage over collaborative safety measures.
• Subcontractor Restrictions Scope: The ban extends to subcontractors, meaning even indirect use of safety-focused AI systems becomes impossible in defense work, creating a comprehensive exclusion mechanism rather than a narrow one.

What Does This Mean for AI Safety Research?

The broader implication is significant for AI safety advocates. If the largest and most powerful customer for AI technology explicitly penalizes companies for maintaining safety guardrails, it creates pressure toward less cautious development. Other companies may conclude that safety features are a liability rather than an asset in competing for government contracts.

Joshua Achiam, an OpenAI researcher, has highlighted philosophical differences between OpenAI and Anthropic on these issues, though Anthropic employees have disputed his characterization. The core disagreement centers on whether AI systems should be designed with hard safety constraints or whether safety should emerge from alignment training and oversight.

The memorandum also directs the Department of Defense to update its AI directive annually, creating a moving target for companies trying to comply. This suggests the government's approach to AI governance will continue evolving, likely in directions that prioritize capability over caution.

For now, the most immediate consequence is clear: companies like Anthropic that refuse to remove safety features face potential removal from the defense and intelligence market. However, the renewable waiver mechanism could indefinitely delay enforcement, leaving the practical impact uncertain. Whether this represents a temporary political decision or a fundamental realignment of U.S. AI policy toward unrestricted development remains to be seen.