The Policy Landscape Is Moving, But Enforcement Hasn’t Caught Up
Washington has responded with a cluster of overlapping initiatives. Executive orders have directed agencies to harden systems against AI-enabled threats on compressed timelines. The General Services Administration has proposed AI Terms of Service to govern how agencies acquire novel AI tools. NIST has published a preliminary Cyber AI Profile mapping AI-specific risks onto its established Cybersecurity Framework, and its Center for AI Standards and Innovation has launched a dedicated effort to standardize security practices for AI agents.
The frameworks are taking shape. The gap is between framework and enforcement—and agencies that wait for a settled policy picture before acting will find themselves behind.
1. DevSecOps Maturity Is the Bottleneck
AI generates more production code than development teams once managed by hand. That speed advantage only holds if the underlying DevSecOps infrastructure can absorb it without compromising quality or security.
In practice, code is reaching production before review processes catch up. Agencies that assume AI output is secure by default are taking on application-security risk they may not be able to see until it surfaces in an audit or an incident. Policy as code and compliance as code are not optional additions—they are the foundation that makes AI-assisted development governable.
2. Agentic AI Introduces Risks That Compound Quickly
Traditional software risks are well-understood. Agentic AI introduces a different category of exposure: agents that have access to sensitive data, can communicate externally, and may encounter untrusted content in the course of their operation.
Two questions sit at the center of that risk. What can an agent actually do with a given dataset? And does that access stay within a single classification level? Without clear answers, agencies are operating with meaningful blind spots.
3. Oversight Must Match the Pace of Output
Governance gaps compound security gaps. Without defined ownership and oversight processes for AI tools themselves, agencies lose the ability to track which agent took which action, under whose approval, and with what access.
The foundation here is not novel—it is the same audit logging, access controls, and change management disciplines agencies already rely on. The difference is that these controls must now be applied earlier, more continuously, and explicitly extended to AI systems. Oversight that lags behind the pace of AI output creates exposure that is difficult to detect before it becomes consequential.
4. Validation Cannot Be Assumed
Agencies cannot expect AI to generate policy-conformant, guidance-compliant outputs without external verification. The speed advantage AI offers is real, but it does not come with built-in compliance.
Treating AI governance as infrastructure—rather than an afterthought—means establishing clear policies for AI use, output validation, and model oversight now, before frontier-model oversight moves from framework to active enforcement.
The Practical Implication
Attackers can act on AI’s expanded capabilities immediately. Defenders’ speed depends on governance that is already in place. That asymmetry is the core problem.
Agencies that treat AI as an addition to established security practices—rather than a replacement for them—are positioned to capture the speed advantage while maintaining the compliance and audit posture their missions require. The ones that pair adoption with real governance, validation, and security discipline now will be better positioned as the rules and the technology continue to develop together.
The ceiling for both attackers and defenders has risen. Which side benefits more depends on how quickly governance catches up.
Comments (0) No comments yet
Want to join this discussion? Login or Register.
No comments yet. Be the first to share your thoughts!