As Artificial Intelligence (AI) continues to unlock new opportunities across the Federal Government, the Office of Management and Budget (OMB) Memorandum M-25-21 provides a framework for scaling innovation while maintaining strong governance, privacy and public trust safeguards. Generative AI (GenAI) is already improving mission outcomes at the U.S. Department of Agriculture (USDA), from citizen services to farm programs and food safety. Zscaler is helping the USDA turn policy into action by providing the visibility, control and governance evidence needed to confidently accelerate AI adoption under M-25-21.
Putting M-25-21 into Practice
Rather than serving solely as a compliance framework, the memo provides a roadmap for scaling responsible AI adoption and maximizing the value of Federal AI investments. Operationalizing the memo centers on six themes:
- AI Use Case Visibility – The foundation for any AI security program.
- Governance and Accountability – Gives the AI governance board evidence to confirm policies work as intended.
- Risk Management – Traditional cybersecurity risks still apply to GenAI, and models introduce additional risk.
- Data Protection – Models must share only intended data with approved models and bots.
- Public Trust – Requires transparency and explainability into how models operate.
- Secure AI Development – Developing secure, private GenAI tools requires a structured methodology.
A Unified Platform for Securing AI
Zscaler supports agency modernization around Trusted Internet Connections (TIC) 3.0 and Zero Trust through a cloud-native, proxy-based architecture that sees end-user traffic and places rich data at agencies’ fingertips. Zscaler’s AI security platform uses AI to counter adversaries by topping cyber threats, preventing data loss and automating security operations. The platform addresses security for AI through three pillars: secure access to AI applications, AI asset management and securing AI applications and infrastructure.
Visibility as the Foundation of AI Governance
The Chief AI Officer (CAIO), governance board and mission owners need a shared view of where AI is used, who uses it, what data it can access and whether it introduces security, privacy or public trust risk. Zscaler’s AI asset management draws on four sources of insight:
- A lightweight endpoint client identifies locally running AI components.
- Zscaler Zero Trust Exchange inspects traffic inline to reveal which users, groups, bureaus and mission areas access which AI tools.
- Code repository scanning shows whether applications invoke models, agents, application programming interfaces (APIs) or Model Context Protocol (MCP) servers before release.
- Cloud environment scanning maps connections among AI assets.
If a developer uses an AI coding assistant to modernize an application, Zscaler identifies the tool and related AI interactions. For a cloud-based chatbot, it also reveals connections to MCP servers, retrieval-augmented generation (RAG) and storage buckets. This visibility surfaces unsanctioned tools, gauges participation in agency-approved GenAI tools and helps leaders understand adoption.
Zscaler then adds risk context by flagging publicly exposed AI service APIs as critical, identifying unnecessary model access as a potential data poisoning risk and assigning model risk ratings for consistent exposure comparison. By mapping each model’s topology, Zscaler detects the data pipeline it can reach and categorizes accessible data, including personally identifiable information (PII) and inspection data, so teams can remove unnecessary connections and limit access to only what each model needs. Supply chain analysis also supports M-25-21’s emphasis on U.S. model provenance and flags known vulnerabilities in dependent models and libraries.
Policy-Based Access to GenAI Tools
Instead of taking an all or nothing approach, agencies can apply policy-based access by user, group, bureau, application and AI model, allowing approved AI use while preventing sensitive data exposure, off-mission use and risky interactions with unapproved tools. Through the Zero Trust Exchange, agencies can allow, block or isolate browser sessions, restrict copy and paste, prevent file uploads and apply watermarks to deter screenshots. These controls support experimentation and productivity while protecting PII, controlled unclassified information (CUI), procurement information, source code and benefits data. Activity is logged to provide auditable evidence for refining policies over time. Agencies can also begin by blocking GenAI entirely, then identify desired tools and build controls around them.
Intent-Aware Guardrails for Prompts and Responses
Traditional data loss prevention relies on keywords, but AI risk often depends on user intent, shared data and model response. Zscaler AI Guard applies trained detectors to understand prompts, as well as malicious, unsafe or off-policy outputs. This intent-based detection can catch threats keyword matching may miss, such as poisoned chatbot data that directs users to malicious links and gives attackers access to USDA systems. AI Guard inspects prompts before submission, blocking or coaching users when sensitive information is involved while keeping data within the USDA and logging activity for review. It can also stop prompt-injection attempts, such as requests for a USDA-built chatbot to ignore instructions and reveal recent inspection findings, before they reach the application.
Building Trustworthy AI Through Continuous Red Teaming
Incorrect AI outputs can impact citizens, producers and the USDA’s credibility. Securing AI development begins with the traditional application lifecycle and Continuous Integration and Continuous Delivery (CI/CD) pipelines, pairing infrastructure security posture insights with runtime controls. AI red teaming completes that picture by testing applications before release, continuously evaluating them in production and turning the results into governance evidence.
Teams test AI applications with preconfigured probes covering security, content safety, hallucinations, business alignment and domain-specific test cases. These probes validate that models do not expose producer or citizen information and avoid making legal, regulatory or policy determinations outside their authority. Because insufficient data can produce hallucinations, each tool must stay within its intended mission and scope. Testing uses more than 5,000 simulated attacks across 70 languages, text and image inputs, multi-turn conversations and encoding methods that conceal instructions directing models to ignore guardrails.
Summary reports explain failures and their causes, helping governance stakeholders adjust controls. Results support remediation through hardened system prompts and policy recommendations for intent-based detectors. Zscaler also benchmarks publicly available models, helping teams select the right foundation before development.
Evidence-Driven AI Governance
M-25-21 requires agencies to establish AI governance bodies, empower accountable AI leaders and develop GenAI guidance, all of which require reporting and evidence. AI governance cannot live solely in a spreadsheet, memo or committee meeting; it requires dashboards and logging. Zscaler helps the USDA align with M-25-21, the NIST AI Risk Management Framework (RMF), the M-22-09 Federal Zero Trust Strategy and best practices such as the Open Worldwide Application Security Project (OWASP) Top 10 for Large Language Models.
AI security is a continuous effort that combines skilled teams, well-defined processes and integrated technologies. Technology deployment is straightforward and well documented, while fine-tuning policy takes time, and Zscaler’s experienced team supports agencies throughout their AI security journey. Equally important is engaging stakeholders across domains early and often. Leaders who pursue AI security alone may encounter resistance or delays, while a team-based approach helps drive lasting success in secure AI adoption.
To learn more about how agencies can operationalize OMB M-25-21 and adopt GenAI securely, watch Zscaler’s webinar, “Meeting OMB M-25-21: How USDA Can Securely Adopt GenAI.”
Carahsoft Technology Corp. is The Trusted Government IT Solutions Provider, supporting Public Sector organizations across Federal, State and Local Government agencies and Education and Healthcare markets. As the Master Government Aggregator for our vendor partners, including Zscaler, we deliver solutions for Geospatial, Cybersecurity, MultiCloud, DevSecOps, Artificial Intelligence, Customer Experience and Engagement, Open Source and more. Working with resellers, systems integrators and consultants, our sales and marketing teams provide industry leading IT products, services and training through hundreds of contract vehicles. Explore the Carahsoft Blog to learn more about the latest trends in Government technology markets and solutions, as well as Carahsoft’s ecosystem of partner thought-leaders.
