Government's shadow AI problem is also a product problem

Vithun Khamsong via Getty Images
COMMENTARY | Cyber rules can block unapproved tools, but cannot erase the incentive to use them when sanctioned AI falls materially behind.
Recent Route Fifty coverage captures where public-sector cybersecurity is heading. Cyber ranges are being used to build trust in AI; the White House and Texas have launched Project Watershed 250 to stress-test water utilities; and the case for whole-of-state cyber defense is that governments cannot protect one agency at a time.
All of those efforts depend on a quieter assumption: public employees will actually use the secure tools their organizations approve.
When an approved artificial intelligence system is slower, less capable, or poorly integrated compared with a public alternative one browser tab away, government creates an incentive for shadow AI. A deadline, a difficult document and a better public tool may be enough. The employee may be trying to help the mission, not evade it.
The federal government has already named part of the problem. In June, National Security Presidential Memorandum 11 directed national-security agencies to close the capability gap between public AI and the tools available to their workforce. Usually read as an adoption policy, it also contains a security lesson: a persistent capability gap is a control weakness.
The Insider Risk May Begin With Good Intentions
Traditional insider-threat programs look for motives such as coercion, grievance, financial pressure, or deliberate rule evasion. Shadow AI can begin with none of them.
Picture a county analyst, state employee, contractor, or utility worker facing a long document and a short deadline. The approved environment can summarize it, but the public model is faster or produces better work. If sensitive information moves with the task, the security failure occurs when the data leaves the governed environment.
That is why punishment and annual awareness training are incomplete answers. Governments can tell employees not to use unapproved tools. They should also ask why the shortcut remains attractive.
A Useful Tool is a Security Control
The State Department offers a useful federal example. Its Generative AI Playbook says StateChat, an enterprise system approved for sensitive-but-unclassified work, had more than 62,000 users by June and had reached 98% of U.S. diplomatic posts.
The scale matters more than the brand. Secure adoption works best when the governed option is useful enough to become the default.
State's own user feedback is equally instructive: employees who reported little or no time savings cited response quality, verification burden and weak workflow integration. Those sound like product issues. In a security program, they are leading indicators. Friction pushes work elsewhere.
A secure AI tool that employees routinely route around is not a successful security deployment. It is a security policy with an escape hatch.
Cyber Ranges Can Measure the Gap, Not Just the Attack
Cyber ranges are already used to rehearse attacks and understand how AI changes cyber operations. The same environments can test whether approved AI is good enough to keep real work inside the boundary.
Give users representative tasks with synthetic or sanitized data and compare the governed system with leading commercial alternatives. Measure completion time, answer quality, verification burden, workflow friction and the steps required to finish the job. Then test the controls around the workflow: what happens when a user tries to move restricted material into an unapproved service, or when a new model materially outperforms the approved one?
That turns shadow AI from a vague policy concern into something an agency can observe before a real incident.
States Can Make This a Shared Service
For state and local governments, the problem is especially practical. A large federal department may have staff to evaluate several frontier models. A small city, school district, water utility, or county health department often does not.
Whole-of-state cyber programs should add shared AI evaluation to the services they provide. A state cyber range or central security office could maintain representative benchmarks, controlled sandboxes, model-risk findings, approved data-use patterns and rapid evaluation procedures that local entities can reuse.
That would not require every jurisdiction to adopt the same model. It would give smaller organizations a way to answer the questions they otherwise struggle to staff: Is this tool materially better? What data can it handle? What controls are required? Can it be approved quickly enough that employees are not pushed toward an ungoverned workaround?
States should also treat product friction as security telemetry. Help-desk requests, blocked uploads to public AI sites, requests for new models, repeated manual workarounds and complaints about response quality can reveal where the control is losing credibility.
Do Not Choose Between Adoption and Control
The wrong response is to frame this as a choice between fast AI adoption and strong cybersecurity. The two reinforce each other when the approved environment is capable, data boundaries are clear and better models can be evaluated at commercial speed.
Technical controls still matter: managed browsers, data-loss prevention, endpoint policy, network telemetry and application-layer controls can block or detect unauthorized transfers. But they work better when compliance does not require employees to accept a large productivity penalty.
The public sector is already learning to stress-test infrastructure and rehearse AI-enabled cyber incidents. It should apply the same discipline to the tools inside the workforce.
The question for the next AI program review should not be only, "Is the approved tool secure?" It should also be: "How far behind is it, who is measuring that distance and what will employees do if the gap gets wider?"
In the AI era, usefulness is not separate from security. It is part of the control.
Burak Oktenli is an independent researcher based in Washington, D.C., focusing on AI assurance, cybersecurity, critical-infrastructure resilience and autonomous-systems governance. He holds a bachelor's degree in Computer Science and Engineering from the University of South Florida and an MBA, and is completing a Master of Professional Studies in Applied Intelligence at Georgetown University.




