Academy
5. Security and Data Privacy
Understand the security guards that inspect generated code before execution, the sandboxed environment, and the guarantee that your data never leaves the platform during processing.
When you ask a natural-language question about your data, OUPI doesn't just hand your dataset to a language model and hope for the best. Instead, the engine writes code from your question, but before that code ever touches your data, it passes through a multi-layered security inspection. The generated code is checked for suspicious patterns — any attempt to access the network, the file system outside your dataset, or the broader platform is blocked and logged. Even the question itself is screened to detect prompt-manipulation attempts. Only after these checks pass does the code proceed to execution. This means no rogue instruction can sneak through to your data environment.
Once the code clears the security inspection, it runs inside a sandboxed environment — an isolated execution space that adds a second wall of protection. The sandbox ensures the code can only interact with the specific dataset you queried, nothing else. Crucially, your data never leaves the OUPI platform during this entire process: the question is interpreted on-platform, the code is generated on-platform, and the computation runs on-platform against your full rows. There is no external round-trip. This is a core guarantee of the architecture, not an optional setting.
Security doesn't stop at code execution. After the sandbox produces a result, the engine writes a plain-language summary — and then a verification step compares every number in that summary against the actual computed data. If any figure doesn't match, a mismatch warning is raised. This post-execution check prevents a model hallucination from slipping into your final answer. Combined with the pre-execution code inspection and the sandbox, you get a three-stage safety net: guard the code, isolate the run, verify the output.
Transparency is part of the security model. Every query keeps its generated code, the result table, the summary, and any suggested chart. You can inspect exactly what code was written and what numbers it produced, so you're never relying on a black box. If the engine misunderstood your question, you'll see it in the code and can rephrase. Saved queries can be re-run on refreshed data and shared with your team, maintaining the same security guarantees each time.
If a query result looks unexpected, click through to the generated code and the result table before questioning the summary. The code shows exactly how your question was interpreted — often the issue is a column name mismatch, not a security or computation problem. Naming columns explicitly in your question ("total revenue by region") reduces ambiguity.
Remember that the engine will not invent missing columns or fabricate data to fill gaps. If your dataset doesn't contain the information needed to answer a question, the query will fail or return incomplete results rather than guess. For complex multi-step analyses, consider using a mission instead.
Open OUPI Data, pick a dataset, and ask a question. When the answer appears, expand the generated code and the result table. Verify that the numbers in the summary match the table. This is the same verification the platform performs automatically — now you've seen it with your own eyes.
OUPI's natural-language query engine protects your data at every stage: generated code is inspected and suspicious patterns are blocked before execution; the code runs in an isolated sandbox with no external access; your data never leaves the platform; and a post-execution verification catches any mismatch between computed results and the summary. Full transparency — visible code, result tables, and flagged warnings — lets you confirm every step yourself.