Safety & Policy Controls
AI governance and guardrails that run on every request. This layer turns your company's rules into controls: who may act, what may pass, when a person must approve, and where work is allowed to run.
What is this layer about
AI systems act on behalf of people and services, so the questions are the familiar ones: who is allowed to do what, with which data, and under whose approval. Safety and policy controls answer those questions in code, at the moment a request is made, instead of in a document nobody reads.
Policies are written once and enforced everywhere: least-privilege permissions on models, tools and datasets, Zero Trust identity for every user and agent, and residency rules that keep data in the regions and jurisdictions you choose. Nothing is trusted because of where it sits on the network.
Around each request sit checks on what goes in and what comes out, human approval for actions that carry real consequences, and isolation for anything that executes code or touches systems. Every decision leaves evidence, so teams can show auditors what was enforced and why.
What do we deliver
Policy enforcement
Business rules become machine-enforced policy on models, tools and data, evaluated on every request and versioned like code, so changes are reviewed and reversible.
Least-privilege permissions
Users, services and agents receive only the access a task needs, for only as long as it needs it. Zero Trust identity is verified on each call, not assumed.
Input and output guardrails
Prompts and responses are checked for personal data, prompt injection and unsafe or off-policy content. Findings are masked, blocked or escalated according to your rules.
Human approval gates
Actions with financial, legal or customer impact pause for a named approver. The request, the context and the decision are recorded alongside the action.
Sandboxing and isolation
Code, tools and agent actions run in constrained environments with limited network, file and credential access, so a mistake or a hostile input stays contained.
Residency and audit evidence
Data residency and sovereignty rules decide where data and processing may live. Each enforcement decision is logged as evidence to support alignment with the frameworks you follow.
Technology stack
- Policy engines
- Open Policy Agent
- Cedar
- Kyverno
- Casbin
- Identity and access
- OpenID Connect
- OAuth 2.0
- SPIFFE and SPIRE
- Keycloak
- Microsoft Entra ID
- Okta
- Guardrails and content safety
- NVIDIA NeMo Guardrails
- Guardrails AI
- Llama Guard
- Microsoft Presidio
- Azure AI Content Safety
- Sandboxing and isolation
- gVisor
- Firecracker
- Kata Containers
- seccomp
- Kubernetes network policies
- Approval and evidence
- Temporal
- Camunda
- HashiCorp Vault
- OpenTelemetry
These are the kinds of technologies this layer is built from; the choice of tools stays with the customer.
Business outcome
Adopt AI inside your risk appetite. Teams move faster when the boundaries are enforced automatically. Leaders can say yes to new use cases because the rules apply the same way every time.
Smaller blast radius when things go wrong. Least privilege, isolation and approval gates limit what a bad prompt, a leaked credential or a faulty agent can reach or change.
Evidence ready for audits and reviews. Decisions are recorded as they are made, which supports alignment with HIPAA, SOC 2, GDPR, DPDP, FedRAMP and PCI DSS without a scramble before each review.
Data stays where it is meant to be. Residency and sovereignty rules keep sensitive data and processing in approved regions, so cross-border questions have a clear, enforced answer.
Frequently asked questions
What does the Safety and Policy Controls layer cover?
It is where your organisation's rules are applied to AI activity: who may use which model, tool or data, what is blocked or sent for approval, and how sensitive content is handled before and after a model responds.
Does this layer make us compliant?
No. The harness helps you align AI use with frameworks such as HIPAA, SOC 2 or GDPR, but alignment is not certification. Compliance still depends on your own policies, processes and the review of your auditors.
Who defines the policies?
Your organisation does. Security, legal and compliance teams decide the rules, and the layer enforces them consistently across applications and models. We can help translate written policy into controls, but we do not set it for you.
How does it work with the layers around it?
It sits between the tools and execution layer below and the evaluation and observability layer above. Controls act on what agents try to do, and the resulting decisions are logged so they can be reviewed and tested.
Questions about your AI rollout?
Talk to us about how the seven layers fit your models and data.