Dropbox details how its in-product AI agent edits files inside isolated VMs with outside permission checks
In an engineering blog post, Dropbox says the agent in its new Dropbox experience grew from two engineers' side projects and runs in per-conversation virtual machines, with credentials and permission checks kept outside the sandbox and conversation-level undo for file changes.
Dropbox has published an engineering account of the AI agent behind its new Dropbox experience, which the company says lets users ask questions about their files and request tasks such as organizing folders, renaming files and preparing materials to share. The post, on the company's dropbox.tech blog, describes the system's design but does not include independent testing. [1]
According to Dropbox, the agent began as two experiments by engineers Daniel Reiter Horn and Rich Chan. Horn built a sandbox with a coding agent and 3D modeling software, connected to Dropbox via the rclone tool, to design a cityscape model for printing. Chan built a graphical agent interface into the Dropbox desktop app to turn receipt images into expense reports without copying files out of Dropbox. [1]
The company says each conversation gets a fresh virtual machine running Ubuntu, launched from a memory snapshot using QEMU through Riviera, its existing content-processing platform that already opens Office documents in VMs to generate previews. Dropbox says the workspace comes online in tens of milliseconds, and that in testing 75% of workspaces were connected to Dropbox and ready to send the agent's first model request within 300 milliseconds of the VM start request. It also says preparation begins once a user has typed more than a few words. [1]
On security, Dropbox says service credentials stay outside the VM, the workspace has no general internet access, and all outgoing requests pass through a proxy to a host process called host_tools that checks destinations, operations and user-granted file permissions before attaching credentials. The company says this means the agent cannot grant itself broader access, and that users see Allow or Deny prompts for folder access. Broad requests require the user to accept a proposed plan first, Dropbox says. [1]
For recovery, Dropbox says it records which conversation produced each file version the agent changes, so users can undo a specific turn without touching unrelated files. If someone else has since edited an affected file, the restored version is saved as a separate, renamed copy, according to the post. [1]
The chat interface is built on the Codex App Server API, and Dropbox says an adapter lets it run Pi, another agent harness, under the same interface and permission controls. [1]
Sources
- Dropbox: How two engineer side projects became the agent taking action inside DropboxSource published · TechDots found it
Statements labelled as a company's claim have not been independently confirmed by TechDots.
Get Technology stories by email
Related coverage
Spotify Engineering questions the idea of a single 'AI product manager' role
A Nov. 9 post on Spotify's engineering blog examines AI product management from three angles, starting from the observation that most coverage treats 'the AI PM' as one job. Only an excerpt was available for review.
Stripe says it moved $1B in vendor payments onto its own Global Payouts product
In a company blog post, Stripe says it began shifting its vendor payments to Global Payouts in early 2026, cutting new-market setup from six weeks to one with a 99.9% payout success rate. The figures are Stripe's own and have not been independently verified.
Stripe says users can now query their Stripe data in natural language through an agent
In a 5 November blog post, Stripe said customers can use natural language with an agent to analyse their Stripe data. Its examples include building an investor slide deck and a custom revenue dashboard. Details beyond the announcement are not available.