Skip to main content

Sandbox hardening profiles

Read this if: you need the concrete containment vocabulary Tyrum exposes to operators and the runtime.

Skip this if: you only need the broader safety model; start with Sandbox and Policy.

Go deeper: Turn Processing and Durable Coordination, Scaling and High Availability.

Profile matrix

ProfileWhat it promisesWhere the limit is
baselineWorkspace boundary checks, sanitized environment, reduced ambient privilegeDoes not assume host-level hardening beyond Tyrum-controlled settings
hardenedAll baseline guarantees plus tighter container/job settings, narrower writable mounts, stricter privilege postureGuarantees depend on runtime support; local subprocess mode cannot over-claim

Purpose

Sandbox profiles give Tyrum a small, explicit containment vocabulary. Operators can choose a known posture, and the runtime can describe realistic constraints instead of guessing.

Deployment semantics

  • In containerized deployments, the profile maps into ToolRunner job or pod settings.
  • In local-subprocess deployments, part of the profile is declarative because host hardening remains operator-controlled.
  • The runtime must report the active posture honestly; a profile label is only meaningful if the documented controls are actually applied.

Observability

  • /status reports sandbox.hardening_profile.
  • Runtime context can surface the active profile so the model does not invent capabilities it does not have.
  • Diagnostics should show when profile-specific restrictions blocked execution.