LLM01 Prompt injection
Content from emails, documents, and web pages is treated as data, not instructions. An agent cannot take a high-impact action because untrusted content asked it to.
HomeAI governance
Security and governance
How we govern the AI agents we build
CloudNSite designs and builds AI agents against the NIST AI Risk Management Framework and the OWASP Top 10 for LLM Applications. Our team brings 25+ years of security and DevOps work, and security is part of how we design an agent from the first scope, not a review at the end. This page shows what that means in practice.
See the NIST mappingNIST AI RMF, function by function
The NIST AI Risk Management Framework organizes AI risk work into four functions. This is how each one shows up in the way we build.
| NIST function | What it asks | How we build |
|---|---|---|
| Govern | Accountability, policies, and roles | Each agent has a named business owner, a written list of allowed actions, and change control for releases. |
| Map | Context, intended use, and impact | The Current State Assessment documents the workflow, the data the agent touches, and what a failure would affect, before any build work. |
| Measure | Testing, evaluation, and monitoring | A defined acceptance test before launch, evaluation suites for operations-scale builds, and monitoring and exception logs after launch. |
| Manage | Responding to and reducing risk | Human approval at named decision points, a defined exception path, and a rollback plan for each release. |
OWASP LLM risks we design against
The OWASP Top 10 for LLM Applications lists the most common ways LLM systems fail. These are the risks that matter most for agents that act inside business systems.
LLM02 Sensitive information disclosure
Agents reach only the data their workflow needs. Redaction, private model deployment, and HIPAA or SOC 2 controls are scoped in where they apply.
LLM03 Supply chain
Models and dependencies are pinned, and an upgrade is reviewed and tested before it reaches production.
LLM06 Excessive agency
Least-privilege access and an explicit list of allowed actions for each agent. Anything outside that list goes to a human.
LLM09 Misinformation
Consequential outputs carry their source references and pass through human sign-off at named points.
LLM10 Unbounded consumption
Rate limits and spend caps on model usage, so a loop or an abusive input cannot run up cost or exhaust capacity.
What this page is, and what it is not
This page describes how we design and build. It is not a certification or a compliance claim. Which controls apply to your workflow, and how each one is evidenced, is set in writing during the Current State Assessment or the AI Initiatives Review.
Bring your security team to the call
Book a free 30-minute AI Strategy Call. Bring your security questionnaire and we will walk through it with the engineers who would build the agent.
Book an AI Strategy Call