Cybersecurity in the Age of AI: Best practices for state and local government

Artificial intelligence (AI) is making cyberattacks on state and local agencies faster to launch and harder to spot. Attackers are using it to copy voices, build fake identities, and find weaknesses in government systems with basic technical skills. Although responsible and safe AI integration can bring efficiencies to government operations, leaders in state and local government can maintain baseline security controls while establishing guidelines for employees and third-party vendors.

How AI is changing cybersecurity threats for government agencies

According to the 2026 NASCIO-Deloitte survey, only 26% of state chief information security officers (CISOs) are extremely or very confident their state’s information assets are protected, down from 48% in 2022. Key emerging threats highlighted by CISOs include AI-powered ransomware services, deepfakes, and automated AI agents searching for system vulnerabilities.

The areas experiencing the greatest rise in threats due to AI are vulnerability exploitation and social engineering. Threat actors use deepfake voices and fraudulent personas to deceive employees within social engineering schemes. Vulnerability exploitation represents the second major area, in which adversaries leverage AI capabilities to identify and breach security gaps within government infrastructure.

CISA cybersecurity best practices

AI-driven attacks still depend on familiar weaknesses, such as weak passwords, outdated software, and employees who follow malicious links. And so, establishing fundamental security measures serves as the primary defense for any government organization.

CISA’s Cybersecurity Awareness Month toolkit shares four core actions:

1. 1

   Educate workforce members on recognizing and avoiding phishing attempts

2. 2

   Enforce complex password requirements

3. 3

   Implement multi-factor authentication, prioritizing phishing-resistant solutions whenever possible

4. 4

   Keep all operational and enterprise software updated

The toolkit also recommends logging, backing up data, and encrypting data. For added protection, it suggests sharing incident information with CISA and moving to a .gov domain if your agency is eligible, to reduce impersonation risk.

CISA’s Cybersecurity Performance Goals 2.0 add two actions for 2026:

• Develop a dedicated incident response strategy for potential threat scenarios, conducting practical simulations annually alongside executive leadership and legal advisors
• Establish operational continuity strategies to recover core systems and sustain vital operations offline, using alternative methods like paper procedures or radio communications

Agencies should also keep an inventory of systems and applications, know where data is and how it’s protected, and keep diagrams of how data moves between systems.

What a government AI acceptable use policy can cover

In addition to CISA’s recommendations, agencies can manage the new risks AI brings: staff using AI tools, code running on local computers, and AI features in vendor products.

Agencies can establish policies governing employee input into AI solutions, such as code or personally identifiable information (PII), to mitigate primary internal operational risks. Establishing an AI policy defines these guidelines and creates a benchmark for holding third-party vendors accountable. Every agency can start with the NIST AI Risk Management Framework, which PayIt’s policies are built on, whether they’re writing their first policy or updating one.

A policy can include:

• What data can go into AI tools, such as code or PII, based on the agency’s risk tolerance
• Which AI models and large language models (LLMs) staff can use
• Whether staff can use personal credentials or only the AI systems the agency designates
• Where a person must review AI’s work before it’s used, whether it’s marketing copy or developing code

To assist with policy enforcement, agencies can additionally deploy monitoring software across web browsers and employee workstations to track AI usage. A primary security risk stems from unreviewed software executing locally on end-user devices. This is a step backward for security because data that once stayed in tightly controlled systems can now spread across the organization, bringing agencies to a familiar risk they dealt with 15 to 20 years ago.

Even if an agency has not formally authorized AI tools, it should not assume employees are refraining from using them. Agencies should keep an inventory of AI systems, deciding which ones are approved and which aren’t. The NIST AI Risk Management Framework recommends maintaining a comprehensive inventory of all AI systems, tailoring management efforts according to associated risk levels.

What agencies can ask vendors in an AI vendor risk assessment

Many of the products agencies already use now include AI features, often added after the agency first reviewed the vendor. Because every added feature alters how a vendor manages agency data, each update introduces new considerations for security evaluations.

Vendors that build AI into their products may train models on agency data or send it to outside AI providers, which creates a data-exposure risk, so agencies can ask these questions:

• What data is used to train your models, and what is your training process?
• Do you use proprietary AI systems, or is our data transmitted to external vendors like Anthropic or OpenAI?
• Which underlying system powers this AI capability?
• Does your AI implementation fall within the scope of your SOC report and independent security audit?
• Which sub-processors or third-party suppliers support this technology?

Cybersecurity Awareness Month serves as a reminder for government agencies to evaluate their security practices. Agencies can strengthen core security using CISA’s toolkit and establish clear policies for AI tool usage. By implementing these proactive measures, agencies can safeguard critical information while still using AI to improve operations.