Divide security responsibility
The shared responsibility model divides platform operations from agent control, data ownership, and decision authority between Duale AI and the customer.
Divide security responsibility between Duale AI platform operations and customer controls across identity, data, model providers, tools, work limits, results, and evidence.
- Pairs each platform commitment with the customer control it enables for review.
- States that greater agent autonomy adds customer controls like narrowed grants and human review.
- Requires multi-factor authentication when a person issues or revokes an agent API token.
- Notes the platform supplies no per-action authorization or human review mechanism by default.
Summaries were generated by AI. Generative AI is experimental.
Duale AI operates the runtime. You own the agent, its data, and every decision the agent’s output feeds. This page states which duties fall where, so a review meeting can start from a shared list.
Use it for planning. It does not vary your contract: the legal documents state the commitments that bind Duale AI.
Duale AI runs the infrastructure; you control access
Duale AI operates and patches the platform, resolves the tenant and agent from each credential, and enforces the policies and grants you set. You keep credentials secret, assign roles, and protect your network, workstations, and build pipeline.
Multi-factor authentication is required when a person issues or revokes an agent API token, or changes their own authentication factors or sessions. Agents and access states what each role can do and how token revocation takes effect.
Each agent control has two sides
Each row pairs one platform control with the decision or configuration you retain.
- Area
- Identity and access
- Duale AI
- Derives the tenant and agent from the credential and enforces your rules
- You
- Keep tokens secret, assign roles, and author the policies and Library grants
- Area
- Data
- Duale AI
- Protects transport and the managed storage it operates
- You
- Choose what you send and protect your own process, network, and activity cache
- Area
- Model providers
- Duale AI
- Calls only providers in your declared pool and ranks routes inside it
- You
- Select provider contracts, retention, region, credentials, and the pool
- Area
- Tools and actions
- Duale AI
- Routes only tools in the agent manifest
- You
- Define each tool, authorize each action, add human review, and make side effects retry-safe
- Area
- Work limits
- Duale AI
- Enforces your deadline and fixed ceilings on delegation and model interactions
- You
- Control admission, concurrency, and total spend
- Area
- Results
- Duale AI
- Enforces the response shape you requested
- You
- Apply domain rules and decide whether the business process accepts the result
- Area
- Evidence and recovery
- Duale AI
- Records part of the platform activity
- You
- Record approvals and exceptions, and compensate or repair an unwanted external action
See Data protection for transport and storage, Authoring tools for delivery, and Access and isolation for grants.
See Plan the deployment for the model pool, Secure integration for approval and recovery, and Evidence and audit for records.
Whatever the deployment shape, you must keep the workload inside the documented scope and limits. The platform supplies no per-action authorization or human review mechanism. If you build neither, the agent acts without them.
More autonomy adds customer controls
Two agents in the same tenant can have different risk because one can read content you do not control, reach important data, or change something outside the platform. Use the threat model to test those three properties together. Narrow a grant, split the work, or put a deterministic check or human decision before an external action.
Return to Security to place this responsibility split in the path of one task.