Skip to main content
Compliance & Governance

Governed by frameworks. Verified by audit trails.

OWASP Agentic AI Top 10 mapped. Privacy Act compliant. Every agent decision logged with full reasoning traces. Compliance is engineered into every agent we build — not retrofitted after deployment.

Frameworks & Standards

What we comply with.

Not a summary of what regulations exist — a mapping of what we actually do about them. Every framework listed here has specific controls implemented and evidence collected.

OWASP Agentic AI Top 10

Every risk mapped to a specific technical control

From 10 December 2026, Privacy Act amendments require any entity using automated decision-making that significantly affects individuals to disclose what personal information is used, what decisions are made by computer programs, and where computer assistance significantly influences outcomes.

This is built into every agent deployment that touches personal data — not an optional add-on.

Execution Architecture

Sandboxed privilege rings.

Agents operate in isolated execution rings inspired by CPU privilege levels. No agent can escalate its own permissions. Every ring has independent policy rules, resource limits, and audit requirements.

R2 ── USER-FACINGR1 ── TOOL EXECR0 ── COREPROTECTED
  1. Ring 2

    User-Facing Output

    Response generation, document creation

    Sandboxed
  2. Ring 1

    Tool Execution

    API calls, database queries, file operations

    Sandboxed
  3. Ring 0

    Core Orchestration

    Agent planning, memory access, prompt assembly

    Protected

Agents cannot escalate from a lower-privilege ring to a higher one. Each ring has independent policy rules, resource limits, and audit requirements.

Agent Safety & Human Oversight

Contained. Controllable. Killable.

Agents operate in sandboxed execution rings inspired by CPU privilege levels. No agent can escalate its own permissions. Every agent has a kill switch. Every high-stakes decision requires human approval.

emergency_home

Kill Switches

Immediate agent termination at any layer. Kill switches are configured per-agent during the build phase and tested before every production deployment.

speed

Circuit Breakers

Automatic throttling when error budgets are exhausted. SLO enforcement prevents degraded agents from continuing to operate. Automatic rollback on policy violation.

undo

Saga Rollback

Multi-step agent operations use saga orchestration. If any step violates a policy rule, the entire operation rolls back automatically — no partial state.

policy

Runtime Policy Engine

Every agent action — tool calls, data access, API requests, output generation — passes through the policy engine before execution. Policies defined in YAML, evaluated in under 0.1ms.

Audit & Traceability

Decision traces, not just logs.

Standard application logs are not sufficient for AI compliance. We capture cognitive decision traces — the full reasoning chain, not just the inputs and outputs. Every intermediate step, tool call, and decision point is recorded.

Captured per agent execution

Most fields are captured for every execution. A few — confidence scoring, outcome tracking — depend on the agent's task and integration surface.

01Timestamp and execution ID
02Input — what triggered the decision
03Context — what data the agent accessed
04Reasoning trace — steps taken, tools called
05Output — the decision or recommendation
06Confidence score
07Model version and prompt version used
08Token usage and latency metrics
09Whether a human reviewed or approved it
10Outcome — if trackable
11Policy evaluation result — which runtime rules fired, pass or fail
12Execution ring level at time of action

Built on OpenTelemetry. Exportable to your SIEM or observability platform — Splunk, Datadog, Azure Monitor, and Google Cloud Logging supported out of the box.

These traces directly satisfy Privacy Act automated decision-making disclosure requirements and provide the evidence base for APRA model risk governance.

Data Handling

Your data stays yours.

We never own client data. We never train on client data. Every byte stays in your cloud account, in your region, under your control.

check_circleClient owns all data — always. We have developer access only, never ownership or persistent access.
check_circleNo training on client data. Cloud provider APIs (Vertex AI, Bedrock, Azure OpenAI) do not use customer data for model training by default.
check_circleAustralian cloud regions by default — GCP australia-southeast1, AWS ap-southeast-2, Azure Australia East.
check_circleData classified in 5 levels: Public, Internal, Confidential, PII/Personal, Privileged — each with specific handling requirements.
check_circleMinimum permissions per task. Service account roles only. All access auditable.
check_circleRetention policies defined per engagement. Client-controlled deletion. No data persists after engagement ends without explicit agreement.

For full technical security controls — encryption, network isolation, authentication, and infrastructure architecture — view our security page.

The next step

Pick a time. Let's talk.

A 15-minute introductory call — no pitch deck, no obligation. We'll tell you straight whether AI agents are the right fit for what you're trying to do.