NewPico selects Xnode Cortx as its Enterprise AI Control Plane, ranked highest by AIMultipleLearn more
Capability 08 · Agent Sandboxes

Run agent code where there's nothing to steal

An agent's loop and the code it writes run in their own short-lived sandbox. The user's token, model keys and tool secrets stay outside, every privileged act is brokered back through Cortx's controls, and outbound traffic is denied unless the run's grant names the host.

Sandbox runsWhere each agent run executed, what it could reach, and proof of the boundary.
Today
41Runs sandboxed
0Secrets inside a sandbox
3Outbound requests denied
What runs where

The work goes in. The secrets stay out.

Each run splits in two. Model-controlled work runs in the sandbox; everything privileged stays with the Cortx control plane and is reached only through the broker.

Inside the sandbox
The agent's loopPlan, act, reflect, for one run
Code the agent writesShell and Python run locally
SkillsUntrusted unless marked otherwise
One run capabilityNames the run, user and tools, then expires
Agent Sandboxes
Kept outside, reached through the broker
The user's tokenNever passed into the sandbox
Model keysModel calls go back through the broker
Tool secretsCredentialed tools run outside
Guardrails and approvalsRun with the secrets, not in the sandbox
The internetOnly hosts named in the run's grant
The one rule

The sandbox holds
nothing worth stealing.

Model-written code never runs where secrets live. The credentials, guardrails and ledger stay with the control plane, and every privileged act is brokered back, so escaping the sandbox gains nothing.

Inside the sandbox, and brokered out
Agent loopinside
Model-written codeinside
Run capabilityinside, expires
Model callsbrokered
Credentialed toolsbrokered
Outbound hostsgranted only

Four steps run for every sandboxed agent run. Select a step to jump to it.

01 · Decide

Each turn, Cortx decides whether the run is off, observed or enforced, and which hosts it may reach, from your regulated profile, signed policy and workspace settings. It mints a short-lived run capability and, if hosts are allowed, a signed egress grant.

02 · Boot

Each run gets its own short-lived pod under a sandbox runtime. Cortx Sentinel checks the guest kernel before the agent starts, and refuses anything weaker than the policy asks for rather than downgrading.

03 · Broker

Shell and Python run inside. Every other tool and every model call goes back through the broker, where the plan gate, guardrails, approvals and audit run next to the secrets.

04 · Tear down

On finish, stop, error or deadline, the run is closed and its pod destroyed. Every credential it held expires just after its deadline, and a reaper ends anything missed.

Three things contain a run, together

A runtime name on its own is a request, not a control. Isolation comes from all three at once.

  • 1 No secrets inside
  • 2 Outbound denied by default
  • 3 A broker that re-checks every privileged act

Every run records where it ran, the tier, the kernel, the image digest and the attestation, and shows it to the user.

Short-lived, empty, closed

Isolation that holds
even when the code misbehaves.

Short-lived

One sandbox per run, gone when it ends

Agent code and skills run in a sandbox made for one run. It has a hard deadline, its one credential expires with it, and teardown always runs.

Credentials stay out

Escaping the sandbox gains nothing

The sandbox gets one run capability and a broker address. Never the user's token, a model key or a tool secret, so there is nothing inside worth taking.

Default-deny network

Outbound traffic denied by default

A sandbox has no route out. Only hosts named in the run's signed grant are reachable, through one proxy, and every grant expires with the run.

Isolation your security team
can sign off on

How a sandbox fails safe, how tools are classified, and where it stands today.

Refuse, never downgrade

If the sandbox or the requested tier isn't available, the run is refused. It never quietly falls back to running next to the secrets.

Unclassified means brokered

Only a short, explicit list of code-execution tools runs inside. Every other tool, including one added tomorrow, is brokered: a wrong guess costs a network hop, not a secret.

Say where it ran

Each run records whether it was sandboxed, observed or refused, with its tier, kernel, image digest and attestation, and shows that to the user.

Where it stands today

Agent Sandboxes ship off by default; you turn them on per workspace, in observe mode first. Recorded runs so far use a gVisor sandbox. The microVM tier is built to refuse to start until its kernel can be attested.

See a run with nothing to steal.

We'll run agent code in a sandbox, try to reach an ungranted host, and show where the run executed and what it could see.