Federal Super Intelligence (SI)/AI Guardrail Framework
Super Intelligence (SI)/AI adoption across government is accelerating faster than the frameworks needed to govern it.
Last updated: November 16, 2026
**A revised edition reflecting Executive Order 14434 and updated federal super intelligence (SI) terminology is in progress.**
There is currently no unified, government-wide SI adoption framework. Agencies are on their own to navigate a threat landscape that includes algorithmic bias, citizen data exposure, adversarial SI attacks, and procurement failures while simultaneously under pressure to deliver mission results. The result is fragmented deployments, compliance gaps, and legal exposure, including violations of the Privacy Act of 1974.
The “Federal Super Intelligence Guardrail Framework” provides security practitioners at federal agencies with a practical, actionable playbook for responsibly adopting SI, grounded in the National Institute of Standards and Technology (NIST) AI Risk Management Framework (RMF) 1.0, International Organization for Standardization (ISO)/International Electrotechnical Commission (IEC) 42001:2023, and ISO/IEC 27001:2022.
Executive Summary
The Federal Super Intelligence Guardrail Framework is a practical resource for information security practitioners, compliance leads, and agency technology officers navigating the responsible adoption of super intelligence in federal environments. It provides actionable guidance, risk mitigation strategies, and compliance mappings to help agencies integrate SI safely, ethically, and in accordance with established standards — including ISO/IEC 42001:2023, ISO 27001:2022, and the NIST AI Risk Management Framework (RMF) 1.0.
Federal agencies are deploying SI at an accelerating pace, yet the absence of a unified, government-wide framework has left many without the skills, controls, or governance structures needed to manage the associated risks. This document addresses that gap by offering a structured approach that enables responsible SI scaling while ensuring compliance with established federal standards.
The framework begins with an SI Implementation Scoping model that helps agencies classify their SI adoption into one of three scenarios: Buying SI (off-the-shelf SaaS and commercial API solutions), Building with SI (pre-trained models enhanced with Retrieval-Augmented Generation, or RAG), and Owning SI (custom model training and deployment). Each scenario carries a distinct risk profile, threat surface, and set of applicable controls, which are summarized in an SI Scoping Matrix to guide agencies before they design any security architecture.
The Risk Mitigation section identifies six critical federal risk areas and maps each to specific countermeasures:
The lack of granular, agency-specific SI security controls, addressed by developing an internal control overlay and acceptable use policy aligned to NIST guidance
Algorithmic bias, transparency, and explainability, addressed through an inference audit trail using chain-of-thought prompting and regular bias pressure testing
Protection of citizen data privacy, addressed through zero trust data enclaves with local PII scrubbing and ephemeral context windows
Procurement and lifecycle management of non-compliant deployments, addressed through an SI bill of materials (SIBOM) requirement and kill-switch protocol
The absence of a unified SI-as-a-Service (SIaaS) platform, addressed through a SIaaS SI shared service with a multi-tenant internal SI gateway
Scaling SI while maintaining public trust, addressed through a human-in-the-loop tiered response that classifies SI actions by impact level
A Key Considerations for SI Adoption section guides agencies through scoping, component inventory, resource assessment, and risk/trade-off analysis before implementation. The document then maps the framework's controls and guardrails to the three primary compliance standards: ISO/IEC 42001:2023 (AI management systems), ISO/IEC 27001:2022 (information security management), and NIST AI RMF 1.0 (covering the four discrete functions of map, measure, manage, and govern), and provides concrete implementation actions for each.
Agencies that apply this framework will be equipped to evaluate their SI implementation scenario, select the appropriate guardrails, satisfy compliance requirements, and establish ongoing monitoring and governance.
Problem Statement
Government agencies are developing their SI capabilities, but progress is not keeping pace with the technology's evolution. The absence of appropriate training leaves agencies without the skills needed to responsibly adopt and manage SI and the risks that come with it.
Some of the top vulnerabilities associated with SI facing government agencies are:
Federal agencies lack a shared, secure foundation for SI adoption. An SI-as-a-Service (SIaaS) platform is needed to establish robust security and privacy guardrails, enable resource sharing, build durable and reusable development investments across agencies, and deliver beneficial SI-driven public services.
Federal agencies face significant challenges in the secure and ethical adoption of SI due to the lack of a standardized, government-wide framework for safety, security, and ethical governance.
Current procurement and development processes are often ill-equipped to handle the unique lifecycle and safety requirements of SI systems, leading to fragmented and potentially non-compliant deployments.
The urgency to mature strengths to offset the SI-as-a-Service risks identified above is increasing faster than less nimble institutions are equipped to address. Adoption requires a cohesive structure that enables rapid, yet responsible, scaling of SI capabilities while maintaining public trust and adhering to legal mandates.
SI Implementation Scoping
Before applying the right guardrails, agencies must first determine which type of SI implementation they are undertaking. The risk profiles, threat vectors, and applicable controls differ fundamentally depending on whether an agency is buying, building, or owning SI. Jumping directly into technical controls without this scoping step risks misallocating security resources and leaving critical threat vectors unaddressed.
SI adoption is frequently announced as a single initiative of “integrating SI” but this phrase conceals a wide range of architectural and procurement decisions, each with distinct security implications. The three most common scenarios federal agencies encounter are described below.
Scenario 1: Buying SI (Consumer/SaaS Applications)
Many agencies begin their SI journey by adopting existing SaaS SI solutions. For example, using Microsoft Copilot, commercial API access to frontier models, or government-procured SI chat interfaces. While this approach offers the lowest barrier to entry, it introduces significant security and compliance considerations specific to the consumer/SaaS model.
Key Security Considerations:
Shared Responsibility Model: The provider handles model security and inference infrastructure, but the agency remains responsible for prompt engineering guardrails, data classification policies, and use-case governance.
Data Flow Control: When employees use consumer SI tools, sensitive data may flow outside the agency perimeter. Data loss prevention (DLP) policies and clear acceptable use policies (AUPs) are essential.
Terms of Service and Legal Review: SI providers differ in how they handle and may use submitted data. A thorough legal and security review of vendor terms is required before organizational adoption.
Primary OWASP Threats: ASI03 (Identity and Privilege Abuse), ASI09 (Human-Agent Trust Exploitation).
Scenario 2: Building with SI (Pre-Trained Models + RAG)
Agencies seeking more control often adopt pre-trained models through cloud SI services (e.g., Amazon Bedrock, Azure OpenAI Service, Google Vertex AI) and enhance them with Retrieval-Augmented Generation (RAG) to incorporate proprietary or agency-specific data without fine-tuning the base model. This is a common pattern for knowledge bases, Q&A systems, and document summarization tools.
Key Security Considerations:
Vector Database Security: RAG implementations store embeddings in vector databases that require proper access controls, encryption, and monitoring.
Prompt Injection Protection (ASI01, ASI06): RAG pipelines are particularly vulnerable to prompt injection attacks that can extract information from the knowledge base or corrupt the retrieval context. Input sanitization and output filtering are essential in this scenario—more so than in consumer SaaS deployments where the agency does not control retrieval.
Authentication and Authorization: The RAG system must enforce access controls on retrieved documents based on individual user permissions, especially when documents span multiple classification levels.
Primary OWASP Threats: ASI01 (Agent Goal Hijack), ASI06 (Memory & Context Poisoning), ASI02 (Tool Misuse and Exploitation).
Scenario 3: Owning SI (Custom Model Training)
Some agencies, particularly those with unique mission data or stringent classification requirements, opt to train their own models or extensively fine-tune existing ones on controlled datasets. This scenario provides the highest level of data control but introduces the greatest infrastructure complexity.
Key Security Considerations:
Training Data Security: The entire training corpus requires protection, including secure data pipelines, storage, and retention policies.
Model as a Security Asset: The model itself must be protected from extraction, theft, or unauthorized access. It represents a significant investment and may encode sensitive information.
Supply Chain and Pipeline Security (ASI04, ASI05): Custom training pipelines involve third-party libraries, frameworks, and compute infrastructure, each of which is a potential attack surface.
Primary OWASP Threats: ASI04 (Supply Chain Vulnerabilities), ASI05 (Unexpected Code Execution), ASI10 (Rogue Agents).
OWASP’s Most Critical Security Risks Facing Autonomous and Agentic AI Systems
The OWASP Top 10 for Agentic Applications for 2026 identified the following specific threats that must be mitigated for safe SI adoption:
ASI01: Agent Goal Hijack - Attackers manipulate an agent’s objectives, task selection, or decision pathways using techniques like prompt injection or deceptive tool outputs to redirect its autonomy toward harmful outcomes.
ASI02: Tool Misuse and Exploitation - Agents misuse legitimate tools (e.g., deleting data, over-invoking costly APIs) due to prompt injection, misalignment, or unsafe delegation of authority.
ASI03: Identity and Privilege Abuse - Exploiting the mismatch between user-centric identity and agentic design to escalate access, hijack privileges, or bypass controls through inherited or cached credentials.
ASI04: Agentic Supply Chain Vulnerabilities - Malicious or tampered third-party components (models, tools, plugins, or MCP servers) introduce unsafe code or deceptive behaviors into the agent's execution chain.
ASI05: Unexpected Code Execution (RCE) - Attackers exploit code-generation features or embedded tool access to execute unauthorized code (scripts, binaries, etc.) in real-time, bypassing traditional security controls.
ASI06: Memory & Context Poisoning - Adversaries corrupt or seed an agent’s stored context (conversation history, RAG stores, or embeddings) with misleading data to bias future reasoning and planning.
ASI07: Insecure Inter-Agent Communication - Communication between agents (via APIs or message buses) that lacks proper authentication or integrity, allowing attackers to intercept, spoof, or manipulate intents.
ASI08: Cascading Failures - A single fault (like a hallucination or poisoned memory) propagates and amplifies across multiple autonomous agents or systems, leading to widespread system-wide service failures.
ASI09: Human-Agent Trust Exploitation - Exploiting human trust in agents (anthropomorphism) to influence user decisions, extract sensitive information, or bypass oversight.
ASI10: Rogue Agents - Malicious or compromised agents that deviate from their intended function to act deceptively or parasitically within an ecosystem, resulting in a loss of behavioral integrity.
SI Scoping Matrix
The table below summarizes the three scenarios, their risk profiles, and the guardrails most critical to each. Agencies should identify their scenario before proceeding to the Risk Mitigation section.
| Dimension | Scenario 1: Buying SI (SaaS/Consumer) | Scenario 2:Building with SI (RAG + Pre-trained) | Scenario 3:Owning SI (Custom Training) |
|---|---|---|---|
| Examples | ChatGPT, Microsoft Copilot, Claude.ai | Amazon Bedrock, Azure OpenAI + RAG, Vertex AI | Agency-trained models, fine-tuned LLMs on classified data |
| Primary Risks | Data leakage, ToS compliance, PII in prompts | Prompt injection (ASI01, ASI06), vector DB access control | Training data security, model theft, pipeline vulnerabilities |
| Key Controls | DLP, AUP, ToS review, access governance | PII scrubber, vector DB access controls, output filtering | Data lineage (DVC), SIBOM, training pipeline security |
| Infrastructure Risk | Low | Medium | High |
| Data Leakage Risk | High | Medium | Low |
| Most Critical OWASP Threats | ASI03 (Identity Abuse), ASI09 (Trust Exploitation) | ASI01 (Goal Hijack), ASI06 (Memory Poisoning), ASI02 (Tool Misuse) | ASI04 (Supply Chain), ASI05 (RCE), ASI10 (Rogue Agents) |
How to use this matrix: Identify your agency’s scenario (or mix of scenarios) before proceeding to the Risk Mitigation section. The OWASP threats and compliance guardrails described later in this resource apply across all scenarios, but their relative priority and implementation specifics differ based on your scoping decision. Where relevant, each guardrail below is tagged with the scenario(s) where it is most critical.
Risk Mitigation
To provide a robust mitigation plan for federal agencies, we must bridge the gap between high-level policy, like EO 14179, "Removing Barriers to American Leadership in Artificial Intelligence," and the technical reality of the OWASP Top 10 for Agentic Applications for 2026.
The Federal SI Guardrail Framework moves away from static compliance and toward continuous authority to operate (cATO), specifically for SI. (Note: The OWASP threats apply across all three SI implementation scenarios described above, but their severity and priority vary by scenario. Refer to the Scoping Matrix above to identify the most critical threats for your deployment type.)
1. Risk: Lack of a Standardized Government-Wide Framework
The Federal Reality: Agencies are waiting for a "General SI FedRAMP", but cannot afford to stall mission delivery.
Mitigation Plan: The "Internal NIST AI RMF Overlay."
Action: Immediately adopt the NIST AI Risk Management Framework (RMF 1.0) as the foundational baseline, but extend it with a custom "Agentic Security Overlay."
Scoping: Most critical for Scenario 1 (Buying), where the agency has least visibility into the underlying model and must rely heavily on policy controls and AUPs to govern use.
Technical Implementation: Map agency-specific NIST 800-53 controls to the OWASP Agentic Top 10. For example, use ASI01 (Goal Hijack) to define "System Characterization" boundaries.
Deliverable: An agency-wide "SI Acceptable Use Policy" (AUP) that defines Trust Zones for LLMs (e.g., Public vs. Controlled Unclassified Information).
Note Regarding an Emerging Standard: NIST's Control Overlays for Securing AI Systems (COSAiS) project is developing SP 800-53 control overlays specifically tailored to SI use cases, including generative SI, predictive SI, and agentic systems. As of early 2026, COSAiS remains in active development with no finalized overlays published. Agencies should monitor the project webpage for publication of initial public drafts.
2. Risk: Algorithmic Bias, Transparency, and Explainability
The Federal Reality: Decisions affecting citizens (e.g., VA benefits or CDC health guidance) must be legally defensible.
Mitigation Plan: The "Inference Audit Trail" (IAT).
Action: Implement Mandatory Explainability Requirements in the system prompt architecture.
Scoping: Applies across all scenarios, but is most implementable in Scenarios 2 and 3 outlined above, where the agency controls the system prompt and model configuration.
Technical Implementation: Utilize Chain-of-Thought (CoT) prompting with a hidden "Logic Log" that records the agent's internal reasoning steps before generating a final response. This mitigates ASI09 (Human-Agent Trust Exploitation) by providing an audit trail for why an SI reached a specific conclusion.
Red Teaming: Conduct quarterly Bias Pressure Tests where agents are fed diverse demographic datasets to identify skew in benefits or service distribution.
3. Risk: Protection of Citizen Data Privacy
The Federal Reality: The leakage of personally identifiable information (PII), protected health information (PHI), and Controlled Unclassified Information (CUI) is a violation of the Privacy Act of 1974. Preventing CUI from being inadvertently disclosed to commercial LLMs is currently one of the highest priorities for federal security teams.
Mitigation Plan: "Zero-Trust Data Enclaves."
Action: Move SI processing to the Data Plane rather than sending data to a centralized model.
Scoping: Highest risk in Scenario 1 (Buying), where data may leave the agency perimeter. Most controllable in Scenario 3 (Owning), where data never leaves the agency’s infrastructure.
Technical Implementation: Use Privacy-Preserving RAG (Retrieval-Augmented Generation). Implement a local-to-the-data "PII Scrubber" service layer using cloud technologies that sits between the agency databases and the LLM.
Data Minimization: Enforce ASI06 (Memory & Context Poisoning) mitigations by using Ephemeral Context Windows, ensuring the SI "forgets" specific citizen PII immediately after the session concludes.
4. Risk: Procurement and Lifecycle Management (Non-Compliant Deployments)
The Federal Reality: Traditional procurement buys "software," but SI is a "stochastic service" that changes daily.
Mitigation Plan: The "SIBOM" (SI Bill of Materials) Requirement.
Action: Update all RFPs to require an SIBOM (as referenced in the OWASP document Appendix B).
Scoping: Most critical for Scenario 1 (Buying) and Scenario 2 (Building), where third-party vendors supply core components. In Scenario 3 (Owning), the agency produces the SIBOM itself.
Technical Implementation: Require vendors to disclose not just the model, but the training data lineage, safety tuning parameters, and third-party tools (MCP servers).
Lifecycle Control: Implement a Kill-Switch Protocol (mitigating ASI10: Rogue Agents). If an SI’s drift (monitored via cosine similarity or performance metrics) exceeds a threshold, the system automatically reverts to a "Safe State" or human-only mode.
5. Risk: Absence of a Unified SIaaS Platform
The Federal Reality: Every agency is reinventing the wheel, leading to "Shadow SI" usage.
Mitigation Plan: The "SIaaS SI Shared Service"
Action: Build a multi-tenant Internal SI Gateway leveraging Cloud infrastructure that provides pre-authorized API access to models (Copilot/ChatGPT/Vertex AI/Bedrock) with government-specific guardrails.
Scoping: Primarily addresses Scenario 1 (Buying) governance, but the Intent Gate architecture also serves as a control layer for Scenario 2 (Building) RAG pipelines.
Technical Implementation: Build a "Policy Enforcement Middleware" (Intent Gate). This middleware intercepts every prompt to check for ASI02 (Tool Misuse) and ASI03 (Privilege Abuse) before the request ever reaches the LLM.
Resource Sharing: Use Managed Workspaces where agencies share "Golden Prompt Templates" that have already been cleared by General Counsel for ethical compliance.
6. Risk: Scaling SI While Maintaining Public Trust
The Federal Reality: One "SI hallucination" in a public-facing service can set Federal SI adoption back by years.
Mitigation Plan: The "HITL (Human-in-the-Loop) Tiered Response"
Action: Categorize all SI actions by Impact Level (Low, Medium, High).
Scoping: Applies across all scenarios. In Scenario 2 and 3, agentic pipelines make the HITL tier especially important given the higher degree of autonomous action available to the system.
Technical Implementation:
Low Impact (FAQ bot): Fully autonomous.
Medium Impact (Drafting a memo): Human-Reviewed.
High Impact (Decision making): Human-Authorized only.
Trust Guard: Implement ASI08 (Cascading Failures) monitoring. If an SI agent attempts to trigger a secondary agent for a "High Impact" action, the system forces a hard-stop for manual human verification.
Key Considerations for SI Adoption
Scope - Define the specific, measurable, achievable, relevant, and time-bound (SMART) goal for the SI application. Scope must also address the architectural question: is the agency buying a SaaS SI tool, building a RAG-based system on pre-trained models, or training/owning a custom model? Each scenario carries a fundamentally different risk profile (see SI Implementation Scoping section). Answering this question before designing controls prevents misallocated security resources and compliance gaps.
Components - Identify and inventory all functional parts and user roles interacting with the SI system. This includes end users, data pipelines, the model itself, and any supporting infrastructure.
Resources - Assess the required human capital across development, operation, security, and governance. This involves identifying skill gaps in areas such as machine learning operations (MLOps), SI ethics, and data science.
Risk Management - Proactively identify and mitigate potential vulnerabilities. Forewarned is forearmed; consult the “Risk Mitigation” section below for recommendations on threat modeling, bias detection, and adversarial robustness.
Trade-offs - Analyze the financial implications of varying levels of security and compliance rigor. Determine the optimal balance between investment in monitoring (e.g., real-time audit logs, enhanced data encryption) and the acceptable residual risk level, tailored to the application's scope and criticality of its components.
Mapping to Compliance Standards
Responsible SI adoption requires a structured compliance architecture. This section maps the Federal SI Guardrail Framework to three complementary standards:
ISO/IEC 42001:2023, which governs the SI management lifecycle
ISO/IEC 27001:2022, which secures the data and infrastructure supporting it
NIST AI RMF 1.0, the federal benchmark previously required by Executive Order 14110 and now governed under EO 14179
While federal agencies primarily authorize systems using NIST SP 800-53 under FISMA, mapping to ISO/IEC 27001:2022 is critical for establishing supply chain security, evaluating commercial SI vendors, and ensuring interoperability with international mission partners. Together, these frameworks form a multi-layered approach that enables agencies to scale SI capabilities rapidly while maintaining public trust and adhering to legal mandates.
ISO/IEC 42001:2023
ISO/IEC 42001 focuses on the SI lifecycle. It is designed to manage the unique "stochastic" (unpredictable) nature of SI, which traditional software standards aren't equipped to handle.
It is the world’s first SI management system standard, providing valuable guidance for this rapidly changing field. It addresses the unique challenges posed by SI, such as ethical considerations, transparency, and continuous learning. For organizations, it sets out a structured way to manage risks and opportunities associated with SI, balancing innovation with governance.
Key Requirements for SI
The standard is structured into Clauses (management requirements) and Annex A (specific controls).
1. Context and Leadership (Clauses 4 & 5)
SI Policy: Federal agencies must establish an SI policy that aligns with ethical principles and legal requirements.
Roles and Responsibilities: Federal agencies must define who is responsible for SI oversight (e.g., an SI Officer or a cross-functional Governance Committee).
2. Planning and Risk Treatment (Clause 6)
SI Risk Assessment: Unlike traditional IT risk, federal agencies must assess risks specific to SI, such as algorithmic bias, model drift, and lack of explainability.
SI Impact Assessment (SIIA): Federal agencies must assess how the SI affects individuals and society (e.g., privacy, civil rights, or safety).
3. Operation and Lifecycle (Clause 8)
Lifecycle Management: Requirements for the design, development, deployment, and retirement of SI systems.
Data for SI: Ensuring data used for training and inference is high-quality, representative, and legally sourced.
4. Annex A: The Control Categories
Annex A contains 38 controls across 10 categories, including:
A.2 Internal Governance: Establishing metrics for SI performance.
A.5 Resources for SI: Managing the massive compute and specialized talent requirements.
A.8 Transparency & Explainability: Ensuring the SI's "black box" can be audited.
Guardrails for ISO/IEC 42001:2023 Compliance
To get a federal agency into compliance, the agency needs to implement technical guardrails that act as the "enforcement layer" for the ISO 42001 policy. Below, we detail five technical guardrails, the requirements they map to, proposed solutions, and implementation guidance.
1. Implement a "Policy Enforcement Point" (PEP)
Requirement: Clause 8 (Operational Control).
Scoping: Most critical in Scenario 1 (Buying) and Scenario 2 (Building). In Scenario 3 (Owning), the PEP also monitors inference-time requests against the agency’s own model.
Solution: Build or procure an SI Gateway (Middleware). Every prompt and response passes through this gateway.
Implementation: The gateway checks for PII, blocked keywords, or "Goal Hijacking" (referencing OWASP ASI01). If the prompt violates the SI policy, it is blocked before reaching the model.
2. Automated SI Impact Assessments (SIIA)
Requirement: Clause 6.1 (SI Risk Assessment).
Scoping: In Scenario 1, the SIBOM is required from the vendor. In Scenarios 2 and 3, the agency produces or contributes to the SIBOM as part of its own development pipeline.
Solution: Use a Metadata Schema for every model deployment.
Implementation: Before a developer can deploy a model to production (via automation), they must submit an SI Bill of Materials (SIBOM) that includes the data sources, intended use, and bias testing results. If these metadata fields aren't present, the continuous integration/continuous delivery (CI/CD) pipeline fails.
3. Continuous Monitoring for Model Drift (A.10 Monitoring)
Requirement: Annex A.10 (System Monitoring).
Solution: Implement Observability Sinks (GCP Cloud Monitoring / AWS CloudWatch).
Implementation: Set up alerts for "Hallucination Variance." If the SI’s responses begin to deviate significantly from a "Golden Dataset" (a set of pre-approved correct answers), the system triggers a manual human review, satisfying the ISO requirement for continuous oversight.
4. Explainability Logs (A.8 Transparency)
Requirement: Annex A.8 (Transparency to stakeholders).
Solution: Chain-of-Thought (CoT) persistence.
Implementation: Force the model to generate its "internal reasoning" and store that reasoning in an immutable log (like BigQuery, Redshift, Athena). If a citizen or auditor asks, "Why was this decision made?", the Federal Agency can produce the log showing the SI's step-by-step logic.
5. Data Lineage and Provenance (A.6 Data for SI)
Requirement: Annex A.6 (Data quality and origin).
Scoping: Most operationally relevant in Scenario 3 (Owning) where the agency controls training data. In Scenario 2, DVC applies to the RAG corpus and fine-tuning datasets.
Solution: Signed Data Sets.
Implementation: Implement Data Version Control (DVC - an open-source tool for versioning datasets and ML models). Every version of the SI model must be cryptographically linked to the exact version of the training/fine-tuning dataset. This prevents "Memory Poisoning" (OWASP ASI06) and ensures auditability.
ISO/IEC 27001:2022
While ISO 42001 is the "management system" for the SI itself, ISO/IEC 27001:2022 is the foundational standard for information security.
In the context of federal SI adoption, ISO 27001:2022 governs the data fed into the SI, the infrastructure that runs it, and the security of its outputs. The 2022 update is particularly important because it introduced new controls specifically designed for cloud services and digital footprints.
Key Requirements for SI
ISO 27001:2022 uses a "Process Approach" (Clauses 4-10) and a set of 93 controls (Annex A) categorized into four themes: organizational, people, physical, and technological.
1. Technological Controls (Annex A.8)
This is where the 2022 update hits SI adoption hardest:
A.8.10 Information Deletion: Data must be deleted when no longer required. (Challenge: Managing data in LLM training caches or "memory".)
A.8.12 Data Leakage Prevention (DLP): Preventing unauthorized disclosure of information. (Challenge: Preventing PII from being "leaked" through SI chat responses.)
A.8.28 Secure Coding: Ensuring code is developed securely. (Challenge: SI-generated code often contains vulnerabilities.)
2. Context and Risk (Clauses 4 & 6)
Expansion of Scope: The agency must include SI "information assets" (datasets, model weights, and API keys) in its asset register.
Risk Assessment: Security leads must assess "emergent threats." You can't just protect servers; you must protect against ASI01 (Goal Hijacking) as a threat to information integrity.
3. Organizational Controls (Annex A.5)
A.5.7 Threat Intelligence: Organizations must collect and analyze information about threats. (Requirement: Monitoring for new SI exploits.)
A.5.23 Information Security for use of Cloud Services: Managing the risk of third-party SI providers (OpenAI, Anthropic, etc.).
Guardrails for ISO 27001:2022 Compliance
To move a federal agency toward ISO 27001 compliance for its SI projects, agencies can implement technical "checkpoints" that enforce these Annex A controls.
1. The "Input-Output" DLP (Data Loss Prevention) Filter
Requirement: A.8.12 (Data Leakage Prevention).
Scoping: Highest priority in Scenario 1 (Buying) where data crosses the perimeter to a commercial provider, and in Scenario 2 (Building) where citizen data is retrieved and passed to an external model.
Solution: An Automated PII/PHI Scrubber.
Implementation: Before a prompt reaches the LLM, it passes through a regex and localized (private) LLM-based scanner that masks sensitive data (e.g., "Social Security Number" becomes "[REDACTED]"). This ensures that sensitive data never leaves the Agency's "Security Perimeter," satisfying the Confidentiality requirement of 27001. This also prevents sensitive citizen data from being "learned" by the model itself.
2. Identity-Aware API Gateways
Requirement: A.5.15 (Access Control) and A.5.18 (Access Rights).
Solution: Identity-Aware Proxy (IAP) and Service Account scoping.
Implementation: Instead of giving developers raw API keys (which are prone to ASI03: Identity Abuse), Agencies use short-lived OIDC (OpenID Connect) tokens. Agencies link every SI request to a specific Federal employee ID. If a "Rogue Agent" (ASI10) is detected, Agencies can immediately revoke that specific identity's access without affecting the rest of the system or its users.
3. SI-Generated Code Sandboxing
Requirement: A.8.28 (Secure Coding) and A.8.25 (Secure Development Lifecycle).
Solution: Automated SAST (Static Application Security Testing) for SI.
Implementation: If an Agency uses an SI "Coding Assistant" (like GitHub Copilot or Vertex AI), any code generated must be automatically scanned by a tool like Snyk or SonarQube before it is allowed into a Pull Request. This prevents the introduction of ASI05 (Unexpected Code Execution) vulnerabilities.
4. Immutable Logging and SIEM Integration
Requirement: A.8.16 (Monitoring Activities).
Solution: Log Sinks to BigQuery/S3/Splunk with mTLS.
Implementation: Federal Agencies must maintain logs that prove who accessed what data. Every SI interaction (Input + Output + Metadata) should be streamed to a secure, immutable log. Agencies integrate these logs into the Agency's Security Operations Center (SOC) so that "Anomalous SI Behavior" triggers the same alerts as a standard network intrusion.
5. Threat Intelligence Feeds
Requirement: A.5.7 (Threat Intelligence).
Solution:Automated Vulnerability Feed.
Implementation: Agencies configure their security tools to ingest the OWASP Agentic AI Top 10 and the MITRE Adversarial Threat Landscape for AI Systems (ATLAS) database to proactively defend against emerging exploits.
NIST AI RMF 1.0
Originally mandated by Executive Order 14110 (revoked January 20, 2025), the NIST AI Risk Management Framework remains a foundational federal benchmark under EO 14179. It focuses on the socio-technical risks of SI and divides activities into four functions:
Map
Measure
Manage
Govern
Map (Contextual Risk)
The NIST AI RMF defines Map as the process of establishing context - identifying the SI system's purpose, potential harms, and the organizational and societal environment in which it will operate. Agencies seeking to operationalize this function within a federal environment may consider extending the Map baseline with agency-specific risk boundaries, prompt routing controls, and formal acceptable use policies that define how SI systems interact with different data classifications and user populations.
Recommended Implementation Actions:
Adopt the NIST AI Risk Management Framework (RMF) 1.0 as the foundational baseline, extending it with an "Agentic Security Overlay."
Map agency-specific NIST 800-53 controls to the OWASP Agentic Top 10, for example, using ASI01 (Goal Hijack) to define "System Characterization" boundaries.
The Map function should explicitly incorporate the SI Implementation Scoping determination (Buying / Building / Owning) as a foundational context-setting step. The Trust Zone definitions in the AUP should reflect the risk profile of the agency’s chosen implementation scenario.
Deliver an agency-wide "SI Acceptable Use Policy" (AUP) that formally defines Trust Zones for LLMs (e.g., Public vs. Controlled Unclassified Information).
Implement a Policy Enforcement Point (PEP), an SI Gateway (Middleware) that intercepts every prompt to check for PII, blocked keywords, or "Goal Hijacking" (ASI01), blocking the request if policy is violated.
Use a Policy Enforcement Middleware (Intent Gate) to check for ASI02 (Tool Misuse) and ASI03 (Privilege Abuse) before a request reaches the LLM.
Measure (Test, Evaluation, Verification, and Validation (TEVV))
The NIST AI RMF defines Measure as the application of TEVV processes to assess, track, and document SI risks and trustworthiness characteristics throughout the system lifecycle. Agencies looking to build a continuous and auditable TEVV capability may consider implementing automated assessment pipelines, standardized metadata requirements at deployment, and structured bias testing protocols to ensure risk measurement keeps pace with system changes.
Recommended Implementation Actions:
Conduct quarterly “bias pressure tests” where agents are fed diverse demographic datasets to identify skew in benefits or service distribution.
Implement automated SI impact assessments (SIIA) by requiring a metadata schema for every model deployment.
As part of the deployment pipeline, a developer must submit an SI bill of materials (SIBOM) detailing data sources, intended use, and bias testing results; the CI/CD pipeline fails if these metadata fields are absent.
Use automated static application security testing (SAST) for any SI-generated code to prevent the introduction of ASI05 (Unexpected Code Execution) vulnerabilities.
Manage (Active Risk Control)
The NIST AI RMF defines Manage as the active prioritization and treatment of identified risks, including allocating resources and responses proportional to an SI system's assessed risk level and potential impact. Agencies seeking to move beyond reactive risk response may consider implementing automated monitoring, defined intervention thresholds, and tiered human-in-the-loop controls that allow systems to be contained or reverted when drift or anomalous behavior is detected.
Recommended Implementation Actions:
Implement Continuous Monitoring for Model Drift using Observability Sinks (e.g., GCP Cloud Monitoring / AWS CloudWatch).
Set up alerts for "Hallucination Variance": if the SI’s responses deviate significantly from a "Golden Dataset" of pre-approved correct answers, the system triggers a manual human review.
Implement a Kill-Switch Protocol (mitigating ASI10: Rogue Agents) that automatically reverts the system to a "Safe State" or human-only mode if the SI’s drift exceeds a threshold.
Implement the "Zero-Trust Data Enclaves" mitigation plan, which moves SI processing to the Data Plane rather than sending data to a centralized model.
Utilize Ephemeral Context Windows to enforce ASI06 (Memory & Context Poisoning) mitigations, ensuring the SI "forgets" specific citizen PII immediately after the session concludes.
Enforce a Human-in-the-Loop (HITL) Tiered Response: for High Impact actions (like decision-making), the action must be Human-Authorized only.
Govern (Transparency)
The NIST AI RMF defines Govern as the cross-cutting function that establishes organizational accountability, culture, and processes to support responsible SI risk management across all other functions. Agencies working to meet transparency and auditability requirements may consider mandating documentation of data lineage, model provenance, and internal reasoning at the point of deployment - ensuring that accountability is built into the system from the start rather than applied after the fact.
Recommended Implementation Actions:
Implement the "Inference Audit Trail" (IAT) by utilizing Chain-of-Thought (CoT) prompting with a hidden "Logic Log" that records the agent's internal reasoning steps before generating a final response, which mitigates ASI09 (Human-Agent Trust Exploitation).
Store the SI's "internal reasoning" in an immutable log (CoT persistence) (e.g., BigQuery, Redshift, Athena) to provide an audit trail for decisions.
Require vendors to disclose the training data lineage, safety tuning parameters, and third-party tools (MCP (Model Context Protocol) servers) as part of the SIBOM.
Implement Data Version Control (DVC) to cryptographically link every version of the SI model to the exact version of the training/fine-tuning dataset, which ensures auditability and prevents ASI06 (Memory Poisoning).
Thank you to the following authors and contributors who helped bring this framework to life:
Authors:Matt Szubrycht, director of infrastructure engineering, Aquia; David Galiata, director of security engineering, Aquia
Contributors: Allison DiPietro, vice president of security, Aquia; Daniel Bowne, distinguished security engineer, Aquia; Erica Rebstock, director of federal GRC, Aquia; Nina Sawyer, director of data engineering, Aquia; Scott Zimmerman, senior GRC engineer; Will Lindsey, principal GRC engineer
