The AI runs its own code to tidy your Excel. Is it safe unreviewed?

The AI's own code does the math and builds the documents
Upload a sales spreadsheet and ask for regional growth rates on a single PowerPoint slide, and the AI can't answer in words. It has to calculate numbers and produce a file. If it does the math by guesswork, one figure comes out wrong, and that one figure makes the whole report untrustworthy. So OwlAI handles calculations and file work with code the AI writes on the spot. Calculations go to a program built to calculate; Word, PowerPoint, Excel, and Hangul documents go to a program built to produce documents.
That cuts the errors that come from guesswork, but the code runs on your company's servers without anyone reviewing it. What makes this acceptable is how it runs. Code the AI writes runs only inside an isolated space built fresh for each execution and cut off from both the internet and your internal network. On a server where that isolation can't be established, it doesn't run at all.
Unrestricted, one faulty run can reach other people's files and the outside world
Most AI-written code does what it was meant to do. Mistakes still happen, though, and the AI can be steered by malicious instructions hidden in an attached document or a web page. Run without restrictions, such code could open other people's files, send data outside, or exhaust the server's resources. Even wrapping it in an ordinary container can leave gaps, depending on how the server is configured.
The code runs only in a space cut off from everything else
The isolated space is built fresh for every run and is connected neither to the internet nor to your internal network. The code has no path for sending data out, and it cannot reach OwlAI's own database. The only files it can see are the attachments in the current conversation and the outputs created in that conversation, so even the same person's other conversations stay out of sight.
Time and resources are capped too. Compute time, process count, and memory for a single run all stay within set limits. System areas are read-only, so the code cannot quietly install programs. When it needs company data, it calls a tool instead of connecting to the database directly, and that call goes through the same permission checks and approval steps as a question from a person.
The most important rule is a single one: if isolation cannot be established, the code does not run. There is no exception mode that runs without isolation, and every refusal to run for lack of isolation is recorded in the audit log.
Whatever the code produces goes out through the same checks. When an Excel or Word file or a calculation result written by code goes to the user, or out to email or an issue tracker, it passes through the same output boundary as a chat answer. Personal-information masking and approval for external sending apply to code output exactly as they do to chat.
Files are stored separately by organization and by person
Files the AI creates and attachments are stored separately for each organization and each person. Someone who belongs to two organizations still cannot open one organization's files from a conversation in the other. When an organization is wound down, its files alone can be selected and deleted.
What your security team should check before adoption
Isolation is built on the security features of the server's operating system, so the same software needs different preparation on different servers. On Ubuntu 24.04 and later, the bundled security profile must be loaded first. The RHEL/Rocky 8 family needs kernel 8.10 or later. If your server hardening standard has disabled the kernel features isolation depends on, every code run is refused.
A check tool verifies these conditions before installation. After installation, the service checks for itself that isolation actually holds and reports the result as a status. SaaS uses the same isolation. When you evaluate OwlAI, tell us the distribution and security policy of the servers you plan to run it on, and we can match the preparation steps precisely.
What technology solves this?
More in Security and audit

Insights · Security and audit
AI is sometimes wrong. Can it handle work involving customer data?
AI evaluations for work usually end on the same question: the performance is clear, but can we really trust it with this?

Insights · Security and audit
You want the newest AI models. The customer data can't leave.
A sales rep pastes a customer list into an AI and asks it to tidy things up.

Insights · Security and audit
A hidden line in an email you asked it to summarize. Will the AI obey?
You ask the AI to summarize twenty emails from business partners.
