AI security and data controls
Port is committed to developing AI responsibly, maintaining high standards of data privacy and security across all our AI interfaces. This page addresses common questions about security, data handling, permissions, and controls for Port's AI capabilities.
Data access & permissions
Port AI strictly respects your organization's existing RBAC (Role-Based Access Control) settings. AI interfaces can only access data that you already have permission to view through Port's standard permissions system. No AI feature bypasses or circumvents your configured access controls.
Port AI only shows data you already have access to through Port's RBAC system. If you see information you didn't expect, it usually means your organization's permissions are broader than intended, not that AI bypassed security controls.
In this case, we recommend reviewing and tightening your RBAC settings to ensure permissions align with your intended access policies.
AI can run tools you have permitted it to use. If you build tools that fetch external data outside of Port and Port AI has permission to run them, it will have access to that external data in those cases. This is a useful way to connect AI to external data sources, but should be used with caution and proper security considerations.
Controlling access to Port AI
Access to Port AI is controlled through the _ai_invocations blueprint's permissions. Users need the Read, Register, and Update permissions on AI invocation entities to access and use Port AI interfaces. By configuring these, you can manage which users in your organization have access to Port AI features, including the Port AI Assistant, AI Chat Widgets, and API-based interactions.
To configure who can access Port AI:
- Navigate to the Data model page in Port.
- Search for the
_ai_invocationsblueprint. - Go to the Permissions tab.
- Under the Entities section, configure the Read, Register, and Update permissions. Only users with all three permissions will have access to Port AI.
A common mistake is granting only the Register permission, or only Register and Read, to members. This will result in a "missing permissions to manage AI invocations" error during use. Make sure to grant Read, Register, and Update permissions so users can view invocations and Port AI can write back results to the invocation record.
Isolate AI records per user
You can grant Members access to Port AI while limiting them to their own records. Apply a dynamic permission policy to both _ai_invocations and _ai_conversation. Each blueprint has an asked_by relation to the user who initiated the interaction or conversation.
-
Go to the data model page and select
_ai_invocations. -
Open Permissions, select Entities, and click Edit JSON.
-
Remove Member from the
rolesarrays forentities.read,entities.register, andentities.update. Keep any administrative access your organization requires. -
In each of those three scopes, set
policyto the following query:{"combinator": "and","rules": [{"property": "asked_by","operator": "=","value": {"context": "user","property": "$identifier"}}]}The query compares the record's
asked_byrelation with the current user's identifier. -
Set
ownedByTeamtofalsefor these scopes and remove any other role, user, or team grants that give the Members you intend to isolate unrestricted access. Click Save. -
Repeat these steps for
_ai_conversation.
Port grants access when any applicable permission check passes. Leaving Member in a scope's roles array gives Members access independently of the policy. Removing the role grant does not disable Port AI when the policies grant the required Read, Register, and Update permissions on each user's own records.
The policy applies to existing records as well as new ones. Records without an asked_by relation matching the current user do not satisfy the policy. AI invocation records cannot be deleted; use permissions to control access to them.
Validate the policy
-
Test with two separate browser sessions signed in as different Members, without administrative or other broad grants. In each session, start a conversation and send a follow-up message to verify that creating, reading, and updating your own AI records still works.
-
Confirm that each user can access their own existing and new conversations and invocations, and cannot access the other user's records.
-
Use the permission simulator on both blueprints to check Read and Update for each user's own records and the other user's records. If access to another user's record succeeds, review the granting path and remove the remaining broad grant.
Organization administrators can:
- Control access to AI features through existing Port RBAC settings.
- Monitor AI usage through audit logs and invocation records.
- Review all AI interactions through the AI invocations catalog.
- Export audit logs for compliance and analysis.
- Configure organization-wide policies for AI feature access.
AI access controls operate at both levels:
- Organization-level: Overall AI feature enablement and policy settings.
- User-level: Individual access permissions based on roles and configured RBAC rules.
Yes. Remove all grants, including dynamic policies, for the Read, Register, and Update permissions on the _ai_invocations blueprint so that no user (including Admins) has access. With these permissions removed, no user in your organization can use Port AI. Some AI elements may still appear in the UI, but they won't function without the required permissions.
You can also contact our support team to opt out at the platform level.
Users cannot individually opt out of AI features that are enabled organization-wide, but administrators can use RBAC controls to restrict access for specific users or roles.
Data privacy & retention
- LLM providers and sub-processors: AI inference data (prompts, tools, and responses) may be retained for up to 90 days.
- Your Port organization: AI invocation records are available as an audit log of AI interactions, tools called, usage, access, cost, and more. These records cannot be deleted.
We use this stored data strictly for:
- Ensuring AI features function correctly.
- Identifying and preventing problematic or inappropriate AI behavior.
- Performance monitoring and system optimization.
- Debugging and troubleshooting issues.
We limit data storage strictly to these operational purposes.
All data processing occurs within Port's secure cloud infrastructure. We use different LLM models depending on the AI interface:
- OpenAI GPT models for certain AI features.
- Claude models for enhanced reasoning capabilities.
Your data is not used for model training by these providers. We ensure complete logical separation between different customers' data throughout the processing pipeline.
Yes, you can contact us to opt out of retaining AI inference data with Port's LLM providers and sub-processors for up to 90 days. Opting out may impact our ability to provide support and troubleshoot issues with AI features. It does not remove AI invocation audit records in your Port organization. Those records cannot be deleted.
Yes. Even when you bring your own LLM provider, AI inference data (prompts, tools, and responses) may still be retained for up to 90 days with Port's LLM providers and sub-processors as needed to run Port AI. This is separate from your LLM provider's policies and from AI invocation audit records in your Port organization, which cannot be deleted. This retention supports:
- Monitoring and troubleshooting AI functionality.
- Error analysis and system optimization.
- Improving the overall AI experience.
This data retention is used to monitor and troubleshoot Port's AI services behavior, prompt engineering, and data processing logic, regardless of which LLM provider you choose. This data is not used for model training.
You can request to opt out of this data retention if required by contacting our support team.
Human oversight & autonomy
AI features in Port operate with different levels of autonomy depending on configuration:
Query responses: AI can autonomously read and analyze your Port data to answer questions. In the Port AI Assistant, ask mode restricts the assistant to read-only tools.
Action execution: You can configure whether AI executes actions automatically or requires human approval before execution. In the side chat, use per-tool approval settings. Via API, use tool approval and executionMode.
For actions with significant impact, we recommend requiring human approval.
- All AI interactions create detailed audit trails.
- Action execution can be configured to require human approval per tool, or at the agent/API level via
executionMode. - Organization administrators can monitor all AI activity.
- Users can review AI reasoning and tool usage for each interaction.
Comprehensive visibility is provided through:
- AI invocation records: Every interaction creates a detailed log.
- Audit trails: Complete history of AI usage across your organization.
- Tool execution logs: See exactly which tools AI used and why.
- Performance metrics: Token usage, response times, and error rates.
- Admin dashboards: Organization-wide usage monitoring.
Rate limits & usage controls
Port AI operates with specific rate limits and monthly quotas to ensure fair usage across all customers. For detailed information about current limits, quotas, monitoring your usage, and what happens when you reach a limit, see Limits and quotas and AI invocations.
These limits apply when using Port's managed AI infrastructure. Bring your own LLM provider to use your provider's limits instead.
Compliance & security standards
Port AI features respect your existing data governance and compliance policies:
- All interactions are logged for audit purposes.
- RBAC controls ensure data access follows your defined policies.
- Data retention policies can be configured to meet your compliance needs.
- Export capabilities support compliance reporting requirements.
For detailed information about Port's security standards, certifications, and compliance frameworks (SOC2, GDPR, ISO 27001, etc.), see Port's Security page.
Troubleshooting & support
AI can make mistakes. If you receive incorrect answers:
- Review the AI invocation details to see which tools were used.
- Check if the AI used incorrect field names or properties.
- Try rephrasing your question or breaking it into smaller queries.
- Provide feedback through the AI invocation record (use the thumbs up/down rating on the AI response).
- Contact support if problems persist.
Remember that AI features are continuously improving, but they are not infallible.
AI features typically start streaming responses within 5 seconds and complete within 30 seconds, depending on:
- Complexity of the query or request.
- Amount of data being analyzed.
- Current system load.
- Which AI interface is being used.
Response times may occasionally be longer as we optimize performance. This is expected behavior and will improve over time.
Response times up to 30 seconds are normal and expected for AI processing. If you experience consistently longer response times:
- Check the AI invocation details for any errors.
- Verify your usage hasn't hit rate limits.
- Contact support if problems persist.
Every AI response includes a thumbs up/down rating. You can optionally add a short comment after rating. Feedback is stored per assistant message inside the invocation's execution logs. For details on how feedback is tracked, see Feedback on AI invocations.
To report issues or share broader feedback, contact our support team.
For any security-related questions or concerns about AI features, contact our support team directly. We take security concerns seriously and will respond promptly to address any issues.