Skip to main content
Compute sandboxes can reach any host by default. Two settings limit one to OneShot’s API plus an allowlist:
  • Per goal: pass egress_allowlist when you create the goal. Setting it, even to [], limits every sandbox of that goal.
  • Per organization: an admin turns on default-deny with PUT /v1/org/sandbox-egress { "default_deny": true }. Every sandbox of its goals is then limited.
Entries are hostnames, IPv4 addresses or IPv4 ranges, at most 50; no wildcards, URLs, IPv6 or all-traffic ranges. The policy is read when a sandbox is created, so it covers queued goals; a running sandbox keeps its policy.

What is enforced

The network filter denies every connection except OneShot’s API and the listed hosts. Its limits:
  • Hostnames are matched on the HTTP Host header (port 80) and the TLS server name (port 443); for other ports, list an address or range.
  • A hostname on a shared CDN endpoint may expose other sites served there.
  • No DNS server is reachable. A resolver inside the sandbox answers only the listed names and refuses every other lookup; a query sent straight to 8.8.8.8 is denied by the filter.
  • An allowlisted address range that contains a public resolver reopens DNS to it, unrecorded: allow rules win over deny rules.
  • Only IPv4 is filtered.

Blocked attempts

Each reported attempt is one egress_blocked goal event and one compute.egress_blocked entry in the organization’s audit log, under Policies: goal, sandbox, destination and time.
  • source: dns_refused: a lookup of a name outside the allowlist, refused and recorded by the resolver whatever program made it and however it exited. The runtime reports it after every tool call and at exit; the agent loop after every command, on pause and on end.
  • source: command_failure: a connection to an address outside the allowlist. The network filter denies it without a log, so it is recorded only when a finished command’s output shows the failure, with the host when named.
A report delivered twice is one record. No command is refused on its text: a URL in a command is data until a connection is made. A root process could stop the resolver; lookups stay denied, but unrecorded. Recording every denied connection is planned.
OneShot tools the sandbox calls (web read, search, browser, email) run outside it and are not limited.

API

Both return { success: true, data: { default_deny } }. Each change is an org.sandbox_egress_changed audit entry.