Security model for Candu Actions

Documentation Index

Fetch the complete documentation index at: https://agents.candu.ai/llms.txt

Use this file to discover all available pages before exploring further.

Candu Actions is built for governed product actions. Your product remains the final execution boundary: Candu manages the conversation layer and structured tool calls, but your app calls your API from the user’s browser using your own credentials and authorization rules. Any action can be configured to require explicit end-user confirmation before it runs. That gives teams a practical control point for sensitive workflows without moving API execution or auth tokens into Candu.

​

Your API stays in control

Candu does not see or store customer auth tokens for your product. Action execution runs client-side inside the customer’s application. Your snippet intercepts the tool call and calls your API from the user’s browser using your own fetchFn and credentials. Candu manages the conversation layer and structured tool calls, but your API execution stays inside your authentication boundary. For sensitive actions, your backend enforces policy before execution. Product permissions, account state, spending thresholds, and business rules stay in your system. See Architecture for the full execution model.

​

Confirmation for sensitive actions

Candu lets teams require end-user confirmation before an action runs. Confirmation can be applied to any action. Many teams use automatic execution for low-risk lookups and require confirmation for actions that create, modify, publish, submit, export, or trigger an external effect. For example, an agent can review a customer account and suggest an update, but the end user can be required to approve the resolved action before it changes customer data. An agent can help prepare a workflow from uploaded information, but submitting or publishing that workflow can require confirmation first. Confirmation gives the end user a final review step before an action reaches your product or another system. Common confirmation-gated actions include:

• Creating or updating a segment
• Publishing an onboarding flow
• Changing account settings
• Updating customer data
• Launching a campaign or message
• Triggering an integration sync
• Exporting data

​

Defined action boundaries

The model does not invent arbitrary API calls. Actions are defined by schema. The model selects from configured actions and fills typed parameters. If an action is not configured, the model cannot execute it through Candu Actions. Your product remains the source of truth for authorization. Your API should enforce its own access controls, validation, and business rules before any action runs.

​

When the AI gets it wrong

Language models can make mistakes, misinterpret intent, or be manipulated by hostile inputs such as prompt injection from a document the agent is reading. Candu Actions assumes this and constrains the consequences.

• Actions are defined by schema, not generated freely. The model selects from your configured actions and fills typed parameters.
• Sensitive actions can require end-user confirmation. The user sees the resolved call before it runs.
• Your API enforces final authorization, validation, and business rules. Candu cannot bypass it.
• Action and conversation history can be reviewed to understand what happened and refine the action or agent definition.

​

Action history and analysis

Candu records action and conversation history for review. Depending on configuration, action history may include:

• Who triggered it
• What action ran
• What inputs were submitted
• What changed or returned
• Whether the action completed, failed, or was cancelled
• Timestamp

Candu also includes agent-assisted analysis to help explain why a particular choice was made and suggest refinements to the action or agent definition. Action history is designed for support, security, and customer escalation workflows. It helps teams review what the agent attempted, what the customer approved, what your API returned, and where a failure occurred.

​

LLM providers and model training

Candu uses LLM providers such as OpenAI and Anthropic where configured. Candu does not use customer data to train foundation models, and Candu does not work with LLM providers or tiers that use customer data for model training. Customers can request Candu’s DPA by emailing security@candu.ai. See Candu’s subprocessor list for the current provider list.

​

User-submitted PII

Users may submit personal data in chats, documents, or action inputs. Candu processes that data under Candu’s DPA to provide the configured feature. Conversations, action inputs, and action outputs may be retained according to workspace configuration and contractual terms. Customers should avoid sending unnecessary sensitive data, design action responses to return the minimum necessary data, and use Candu’s deletion process for deletion requests.

​

Optional Code Chat

Some Candu AI features, including Code Chat, require explicit customer enablement. When GitHub-connected Code Chat is enabled, Candu accesses only the repositories selected by the customer using the configured GitHub integration. Under Candu’s DPA, customer code is processed to provide the feature and is not used to train third-party models. Candu’s DPA states that Candu does not retain customer repositories or create persistent indexes or embeddings of customer code for Codebase Chat. This is separate from product action execution. Candu Actions can execute governed product actions without requiring GitHub access. See Connect GitHub to Candu Code Chat for setup details.

​

Data retention

Action history and chat traces are retained according to workspace configuration and contractual terms. Action history may include user messages, resolved action inputs, API responses, completion status, and timestamps. Sensitive inputs can be deleted on request under Candu’s data deletion process, subject to legal, security, backup, and audit-log retention requirements described in applicable agreements.

​

Data processing and subprocessors

Candu processes customer data as described in its data processing, security, and subscription terms. Useful review links:

• Candu Security page
• Subscription Agreement
• Candu subprocessor list

To request Candu’s DPA or additional security materials, email security@candu.ai.

​

Contact

Need a security review? Email security@candu.ai.