For the complete documentation index, see llms.txt. This page is also available as Markdown.

Security

Aura was built with isolation, read-only access, and sandboxing as foundational design principles.

Data Isolation

Read-only guarantee

There is no write path from the AI agents back to the source data.

Constraints on what Aura can do

  • No production database connection. Aura holds no connection string or credentials for the live database, and has no network path to it. It only processes pre-exported, static snapshots.

  • No data modification. The read-only database configuration and immutable format mean there is no write path to any data source.

  • No persistence between runs. Each analysis cycle uses a fresh snapshot. The container environment is disposable, and analysis artefacts are discarded after reports are generated.

  • No cross-tenant access. Each tenant deployment is scoped to its own data partition, with separate storage paths and access credentials. There is no shared data layer between tenants.

  • No arbitrary code execution on customer infrastructure. AI agents run inside an isolated container within the Aura environment, with no access to the customer's network, servers, or systems.

Network Isolation

The Aura container enforces a network allowlist at the firewall level. This is a deny-by-default configuration: any destination not explicitly listed is blocked.

Containerised Sandboxing

Aura's execution environment uses layered isolation with private subnet, no public IP, runs as a non-root user with limited capabilities in its Docker Container and the filesteme restricted to specific AI agent processes.

AI Model Interaction

Data sent to the model

When an analysis cycle runs, the following categories of information may be sent to the LLM:

  • Aggregated metrics — statistical summaries such as OEE percentages, downtime totals, and production counts, rather than raw row-level records.

  • Query results — SQL result sets, limited to the columns and rows relevant to a specific hypothesis being tested.

  • Analysis prompts — instructions describing what to investigate, referencing the customer's facility context and KPI definitions.

  • Operator comments — free-text observations from the MES event log, used to corroborate statistical findings.

Safeguards

  • Encryption in transit.

  • No model training on customer data.

  • Regional data residency.

  • No cross-session persistence.

For technical questions about this architecture, contact your Roima account team. See here.

Last updated