Enterprise-grade security and controls for your environment
Our default deployment runs in your infrastructure, for maximum security and privacy. Access permissions, model connections and approval controls are configurable to meet your security requirements.
What security teams ask us.
Details on deployment, data handling and agent controls.
Have a question that isn't answered here?
security@yantratech.com
01Your environment and your data
Where it runs and what leaves
By default, the application runs in your infrastructure, under your control. This deployment does not require a Yantra-hosted service. You control the infrastructure and access. Model requests are handled by the providers you select, as described below.
Model requests go to your approved model providers. The application uses API keys you provide for your chosen model providers, so your data stays within your own agreement with each provider. Your agreement and account settings with each provider govern how it handles your data. Your integrations, execution services and monitoring configuration determine other external data flows. Connections use your approved accounts and access methods. Monitoring information excludes business data and identifying details. Processing locations are checked against your requirements and the selected service and model configuration before deployment.
No. Yantra does not use your data, generated outputs or client-specific agent memory to train or fine-tune models. Your data stays yours. You choose and control your model-provider setup.
Who and what gets access
The application uses your identity provider and permissions model. You manage users internally and decide who can access what. Within the application, you can also set granular permissions that control what each user can see.
Application service accounts can access only the mailboxes, file locations and databases you authorise for your deployment. Database access used to gather evidence is read-only. These accounts are separate from user accounts and Yantra staff accounts.
Yantra staff use client-issued accounts and access methods you approve. Your policies apply, and you can revoke access at any time.
Data, credentials and updates
The application follows your data-handling and encryption rules and supports your key-management services. You specify the requirements for storage, backups, retention, export and deletion.
Credentials are stored in the environment you approve. Your deployment plan identifies who issues, rotates and revokes them.
Only with your approval. Permitted data use is defined for each environment. A test or proof-of-concept label does not itself authorise the use of production data.
Updates follow your deployment and change-approval process. Testing, maintenance windows and recovery arrangements are defined before changes are applied. Code changes are reviewed, tested and scanned for security issues before release.
02Agent controls, records and recovery
Approval and limits
No. In this deployment model, agents investigate and propose resolutions. A human operator reviews the evidence and decides what happens next. Application code executes business actions only after approval from a named user is recorded.
You control the available tools, accounts and permissions. Read-only database connections have limits on simultaneous queries and running time. Each run has a token budget and limits on creating additional agents. Application controls enforce these limits outside the model. Model responses cannot grant extra access or override the limits.
This is known as prompt injection. The application treats retrieved content as untrusted evidence. That content cannot change permissions or bypass approval: tool and permission limits are enforced outside the model, and business actions require approval from a named user. Injected text can still mislead the model’s analysis, so each proposal shows its source records for your team to check before approval. Protection does not rely on filtering alone.
Checks and records
The application shows the proposed resolution together with the agent’s reasoning, the rules it applied and the source records, such as original emails and trade records. Your team checks the evidence before approving an action. Model outputs can be incorrect or incomplete. The proposal, evidence and human decision are recorded.
Your deployment specifies how failed requests and invalid outputs are reported, when retries are permitted and when a person must investigate. Retry, fallback and escalation behaviour is defined for each deployed workflow.
Agent logs record investigation steps, turns, tool calls, model responses, proposals and stated rationale. Application records capture human approvals, overrides, users and timestamps. Record coverage can be mapped to your regulatory requirements. Any required protection against alteration must be specified and verified for your deployment.
Prompts, responses and investigation logs can contain business data. By default, application logs containing client data stay in your environment. Access, retention, redaction and export requirements are defined for your deployment.
Changes, stops and recovery
Agent configurations are versioned so you can return to previous settings. Shadow runs compare configurations before changes reach the live setup. Model, prompt and tool changes follow your testing and approval process. Provider changes are checked separately. Returning to previous settings does not guarantee that an earlier model is still available.
Your team can stop the application directly in your environment, without waiting for Yantra. To disable individual connections, revoke the relevant model-provider, database or mailbox credentials or permissions. Disabling the application in your identity provider restricts sign-in. These measures do not undo completed actions.
Recovery procedures cover work in progress and queued work, checks before restarting, and how your team continues manually. Recovery and support commitments are defined for your deployment.