The War Department owes Congress a report by Aug. 31 that could shape how the Pentagon buys and secures artificial intelligence for years. Section 1512 of the fiscal 2026 defense authorization law requires a comprehensive review of how the department protects AI and machine-learning systems, covering current practices, gaps, commercial runtime-security options, and alignment with industry frameworks.
The deadline arrives as the administration has set a clear direction on AI security. President Donald Trump signed Executive Order 14409 on June 2 to accelerate AI-enabled cyber defense and rejected the idea that security requires a new licensing regime for AI development. On July 13, the Pentagon suspended Phase II of the Cybersecurity Maturity Model Certification (CMMC) program and launched a review aimed at reducing compliance burden while preserving cybersecurity.
The question is whether the Aug. 31 report matches that standard. A useful review will identify which AI-security controls actually change operational risk, which can be measured continuously, and which should become enforceable contract terms. The risk is that the review translates new technology into familiar compliance language. Program managers would get another checklist that says little about what happens when a model is manipulated at runtime.
What Congress asked for
Congress already structured the next step. Section 1513 of the same law requires a risk-based security framework for AI systems acquired by the Pentagon. It directs the department to cover supply chain risks, data poisoning, adversarial tampering, unintended exposure, continuous monitoring, and incident reporting. It also instructs the Pentagon to extend or augment existing frameworks, including CMMC, and to amend the Defense Federal Acquisition Regulation Supplement or take similar action so covered contractors must implement the resulting practices.
That same section contains quietly important limits. The rules must be narrowly tailored, calibrated to the specific AI system and the contractor's role, and designed with the cost of slowing development in mind. The suspension of CMMC Phase II gives the Pentagon a chance to avoid a certification bureaucracy around AI while still imposing requirements that can be tested and audited.
Security claims need evidence
For AI systems, the new at requirements should focus on evidence. A contractor developing or hosting a model for defense use should be able to show which data and system artifacts are protected, who can change them, how unusual access is detected, what happens when a system receives manipulated input, whether runtime an armed conflict… Whether runtime controls can restrict an unsafe action? and which telemetry survives for inspection. The standard should separate a security claim from proof that the control actually worked in a test.
That's legally and financially consequential. The Justice Department already uses the False Claims Act against government contractors that knowingly misrepresent cybersecurity compliance. In June, defense contractor Logzone agreed to pay over $500,000 to resolve allegations that it knowingly failed to comply with Navy cybersecurity requirements. Earlier cases have led to multimillion-dollar settlements. Once AI-specific cybersecurity duties become contractual requirements, knowingly false looks like the same liability.
This does not make every defense failure a fraud. The False Claims Act has a specific knowledge-and-materiality structure. But contractors should expect their AI-security assertions to become enforceable commitments, not just slide decks and technical white papers.
The building blocks
The Pentagon is already building the institutional machinery. The fiscal 2026 law required the department's AI sandbox task force to brief congressional defense committees by Aug. 1 on its goals and objectives. Trump's June peace on AI national security memorandum emphasized robustness that remains and controllability, accountability, incident response, and secure access to advanced models. The efforts converge on a single question: how does the government know that the planned security properties of an AI system survive deployment?
The Aug. 31 report should answer that question in operational terms. Congress should expect a short list of controls that can be demonstrated, monitored, and audited; a plan for converting the highest-value controls into procurement requirements; and a way to scale obligations with the sensitivity of the model and mission.
What this means for legal professionals
Legal teams advising defense contractors to need to track this closely. If the Pentagon turns Section 1513 into contract language-and the law directs it to-AI-security performance claims will sit behind certifications, invoices, and payment requests that the government can later examine under the False Claims Act. The administration is right to resist regulation that slows AI without buying real security. The defense acquisition should avoid another compliance layer that measures paperwork. The question is whether contractors have the evidence to back their AI-security pictures.
Your membership also unlocks:
