Agent security
How the Sazabi agent investigates on your behalf with least privilege — running commands in an isolated, ephemeral sandbox with credentials it only holds at runtime.
The Sazabi agent reads your telemetry and runs commands to investigate problems. It is designed to do that with the least access needed, in an environment that cannot reach Sazabi's internal systems or persist your credentials.
The sandbox is isolated and ephemeral
When the agent needs to run a command-line tool, it does so in a sandbox — an isolated environment created fresh for a thread and torn down afterward. It starts clean, runs only the commands the investigation needs, and does not persist between threads. Each thread gets its own sandbox, and nothing you connect leaks from one to another.
The sandbox has outbound internet access so tools can reach the services you connect it to, but it is isolated from Sazabi's internal network. The agent's commands run only there — model-authored code never runs inside Sazabi's core services.
Credentials live only as long as the sandbox
Credentials you connect for the sandbox — through a sandbox CLI or a custom CLI — are encrypted at rest and materialized inside the sandbox for the life of that thread's sandbox: commands the agent runs during the thread can use them, and they are destroyed with the sandbox rather than persisting between threads. Because the agent's commands can read them while the sandbox lives, scope them tightly: prefer read-only or least-privilege credentials, and avoid giving the agent production write access. See Set up the sandbox for how connections and credential testing work.
Investigations stay scoped to your project
The agent queries only the data in the project you are working in, and it works through the same access boundaries as everyone else — it cannot reach another organization's data. When it answers, it shows the queries it ran, so you can verify its reasoning rather than take a conclusion on trust.
Your data trains no models
Running an investigation does not expose your data to model training. As covered in Data handling, Sazabi uses enterprise API tiers that contractually exclude your content from training, and passes that commitment through in its Data Processing Addendum.