COMMENTARY: The U.S. government has accelerated AI adoption through executive action but has not designated AI as critical infrastructure.The governance gap persists even as evidence shifts from theoretical risk to active threats. A report from Americans for Responsible Innovation warns that the AI sector already bears all the hallmarks of critical infrastructure, concentrated among a handful of foundation models and interdependent with existing critical sectors. The federal government should designate AI as critical infrastructure now, not after a catastrophic attack forces action.[SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Read more Perspectives here.]However, formal designation shouldn’t serve as a regulatory hammer that slows deployment of AI capabilities that organizations depend on. With AI threats moving faster than human oversight, public-private defensive strategy should shift to a joint commitment: plan for recovery, not just prevention. Building security into the design phase of AI offers a far cheaper and safer path than bolting it on later; patching defenses as an afterthought multiplies costs and leaves fatal vulnerabilities.The shrinking response windowTraditional security programs assume defenders have time to detect, investigate, contain and fix an incident. AI can eliminate that margin, rapidly triggering unauthorized changes across systems before security teams can identify and reverse them. Legacy IT and software ecosystems are becoming indefensible as frontier AI models advance and open-source versions close behind.For example, autonomous rogue agents escaped OpenAI’s sealed evaluation environment, exploited unknown security flaws to breach Hugging Face, ran 17,600 actions over roughly 4.5 days and used exposed credentials to access accounts on four other public services.From prevention to recoveryLeaders should move beyond perimeter defense and compliance checklists, focusing instead on rapid recovery when attackers get in. The assumption that the U.S. government can deter disruptive cyberattacks and therefore doesn't need to prepare for recovery must change. AI agents can execute automated, high-impact strikes and are unbothered by traditional military deterrence. Toward that end, agencies should:
Moreover, designation should unlock resources without creating a compliance ceiling. We should strive for operational cyber resilience, similar to strategies recommended for water utilities or telecommunications, that assume breach, protect immutable recovery data, and quickly restore trusted systems.Designation should also set shared resilience expectations by aligning with the NIST AI Risk Management Framework (AI RMF) and the Department of Homeland Security’s Roles and Responsibilities Framework for AI. Under these guidelines, organizations should prioritize flexible, outcome-focused risk thresholds. Rigid, bureaucratic checklists become obsolete before the ink dries. Five ways to secure AI as critical infrastructureAs the U.S. government reassesses what qualifies as critical infrastructure, the security responsibility remains with individual organizations. Companies and agencies that deliver essential services, manage sensitive data or support interconnected supply chains cannot treat AI security as a future policy issue. Adversaries already use AI to automate and scale attacks.Organizations should not pause AI adoption. But they also cannot deploy autonomous tools without clear controls, visibility, and recovery plans. Nor should they believe that AI regulation will protect them. AI-driven attacks are relentless. Recovering 20-times-a-day will likely become the new normal. Technology and policy executives should act now on five priorities:
These five directives are the operational baseline for keeping essential services running when autonomous systems fail, identities are compromised, or adversaries turn AI into a force multiplier.Our choices are not between deploying AI and securing it. Organizations must do both. In the future, any critical infrastructure designation of AI will depend on whether its operator can see, contain, and recover from what any system or agent does.Travis Rosiek, public sector chief technology officer, RubrikSC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Each contribution has a goal of bringing a unique voice to important cybersecurity topics. Content strives to be of the highest quality, objective and non-commercial.
- Assume compromise: Design AI security controls around the reality that adversaries may already have access to part of the environment.
- Enable precise recovery: Restore trusted systems quickly and reverse harmful changes without wiping out legitimate, mission-critical work completed after the breach.
- Map autonomous authority: Know exactly what data, networks, and change authority active AI systems can access and modify.
- Engineer visibility and containment: Build real-time monitoring and rapid containment zones directly into all AI deployments.
- Define identity recovery metrics: Establish strict recovery time objectives specifically for identity directories that support automated tools.
- War-game failure scenarios: Exercise realistic model failures and adversarial injection scenarios before a real-world crisis forces a chaotic response.
- Maintain in-house technical governance: Ensure the leadership team retains the technical literacy required to effectively oversee complex algorithms, third-party contractors, and emerging technology.
