The White House signed an Executive Order this morning directing federal agencies, state and local governments, and critical infrastructure operators to prioritize AI cybersecurity — most of it on a 30-day clock. Cyber-focused agencies have some new mandates, including standing up a new AI cybersecurity clearinghouse.
While all of that spins up, somewhere in a government building a contracting officer has just uploaded a vendor contract into the latest AI tool and asked it to compare terms against a previously approved agreement, to flag any risks and highlight the gaps. It may be a relevant use case, but it’s highly likely it’s an unmanaged one. Nobody approved that AI session. Nobody can see what might have leaked outside the enterprise.
And though the EO focuses on leveraging new and novel AI capabilities for cybersecurity, the way AI is used every day by government workers was notably absent.
What the EO on “Promoting Advanced Artificial Intelligence “Innovation and Security” Covers
The order directs CISA to release Binding Operational Directives to harden civilian federal information systems, expand AI-enabled defensive tools, and facilitate security tool access for state and local authorities and named critical infrastructure operators (rural hospitals, community banks, and local utilities.)
It also mandates the creation of a voluntary clearinghouse to coordinate AI-discovered vulnerability scanning across industry and government. AI companies are asked to share frontier models with government reviewers before deployment, on a voluntary basis, with a 30-day window.
There is no mention anywhere in the order — or its factsheet — of workforce AI use, shadow AI, or AI data leakage.
This EO focuses on what AI can do to help secure government systems. That’s not a bad thing. But it had nothing to say about what government employees do with AI.
What Our Shadow AI Research Shows: Three Gaps That Directives Don't Reach
Neon Cyber commissioned an anonymous survey in May 2026 of more than 225 knowledge workers across a variety of industries. There were approximately 13 government or public sector respondents. We know this isn’t a statistically representative slice of the federal workforce. We're not presenting this research as definitive estimates. But the differentials we highlight below between government respondents and the broader sample point at something structural.
The policy vacuum: 23% vs. 63%
Only 23% of government respondents said they have a clear AI policy they understand. Across all industries, that figure is 63% — a 40-point gap. Nearly 70% of government respondents fall into what we're calling the policy gap category: policy is unclear, absent, or they've simply never seen it, at more than double the all-sample rate.
A GAO report published in July 2025 found that generative AI use cases across 11 federal agencies had increased nearly nine-fold in a single year, from 32 in 2023 to 282 in 2024. The same report found that officials at 10 of 12 agencies identified existing federal policy as a potential obstacle to adoption.
Meanwhile, OMB's April 2025 memorandum M-25-21 explicitly directs agencies to establish clear expectations for their workforce on appropriate AI use — and states that every federal worker has a responsibility to develop and maintain, at minimum, foundational knowledge of how to use AI responsibly in performing their official duties. Our survey data highlights that workers are still lacking in clarity: less than a quarter say they have a clear policy they can actually act on.
The procurement waiting room: 69% vs. 41%
Government respondents were nearly twice as likely as the all-industry sample to expect approval for a new AI tool to take more than two weeks — 69% versus 41%. Several expected that approval to take months.
The EO calls for rapid deployment of AI-enabled defensive tooling. Procurement reality creates a waiting room.
What's notable in our data is what workers do in that gap: government respondents actually show lower workaround rates than the broader sample. They're more likely to stop using AI or wait for approval rather than route around controls.
In this regard, unlike private sector enterprises, the problem isn't the exploding use of shadow AI tools, but — as has often been the case in the past — a slower adoption of innovative technology that could help the government improve efficiency.
The browser identity gap and where data is leaking: 92% vs. 69%
92% of government respondents reported using the same browser profile for personal and work activity, versus 69% across all industries. Professional identity and personal identity are running in the same session, with no separation between them — which means an AI interaction, including ones involving credentials or internal government documents, is happening in an environment that looks identical to a personal browsing session from any monitoring tool's perspective.
The EO focuses on leveraging AI for vulnerability scanning and to harden systems. Vulnerability exploitation is serious, but it takes some effort. You know what doesn’t? A government worker uploading an internal report to Claude.
46% of government respondents reported sharing internal business information — strategy documents, emails, presentations, meeting notes— with an AI tool, versus 31% across all industries. And 38.5% of government respondents reported pasting sensitive data into AI tools most or almost every time — in line with the 52% all-industry rate, meaning this is a workforce-wide pattern playing out in an environment with far less policy cover.
Even more shocking: Of the government respondents in our survey, 23% self-reported pasting login credentials or API keys into an AI tool. Across all respondents, that figure was 11%. In a sector where credential access can mean compromised access to federal systems, sensitive citizen data, or critical infrastructure controls, that's not a throwaway data point.
The Layers the EO Doesn't Reach
The EO addresses the systems: the network, endpoints, traditional IT infrastructure, even the AI model itself. But it doesn’t address what employees are doing with AI inside the browser session.
Network controls see traffic flows and known malicious destinations — not the content of an HTTPS session to a legitimate AI provider. Endpoint security sees the device, not what the user pasted into a prompt or which contract they just uploaded to Claude. Identity providers see who authenticated to sanctioned applications, not the Gemini session running in the same browser profile as the worker's personal Gmail.
The EO was written in response to a real anxiety: that AI can find and exploit software vulnerabilities at a speed and scale no human team can match. That's the model-layer threat — AI as an attacker tool.
The unaddressed threat is at the workforce layer: AI as an inadvertent exfiltration surface, where employees are feeding sensitive data into sessions nobody can see, not because they're malicious but because it's the fastest path to getting work done.
The EO is focused on hardening one side of the problem. The workforce side is still wide open.
Our data underscores that government workers aren't the variable here. Despite operating with significantly less policy clarity, longer approval timelines, and a blended personal and professional browser environment, their compliance behaviors are broadly in line with the wider workforce — but so are the risk outcomes.
Want Early Access to the Research?
We've been analyzing the research. The full findings from our 2026 shadow AI survey are coming shortly. We’ll cover behavioral patterns across industries, the kinds of data moving through unmanaged sessions, and what the organizations with the least exposure are doing differently.