maf-sandbox-acas

Renamed. This package was published as maf-sandbox-aca up to 0.2.3. ACA is Azure Container Apps, the broad service, while this backend targets Azure Container Apps Sandboxes — so it gained the s the rest of the project already used. PyPI names cannot be reused, so this is a new distribution rather than a continuation, and there is no compatibility shim: maf_sandbox_aca and the Aca… classes do not forward here. See #68.

Experimental. This package is early-stage (pre-1.0, Development Status :: 4 - Beta) — its API may change or be removed in a future release without notice. Importing it emits a one-time MafSandboxAcasExperimentalWarning; suppress it with warnings.filterwarnings("ignore", category=maf_sandbox_acas.MafSandboxAcasExperimentalWarning) once you've read the notice.

This package is not affiliated with, endorsed by, or a product of Microsoft — it is a third-party reference implementation of microsoft/agent-framework#7568 for Microsoft Agent Framework, built on the Azure Container Apps Sandboxes preview.

app  ->  maf_sandbox  ->  maf_sandbox_acas  ->  the sandbox

An agent that writes code should not be the thing that runs it. This package gives it somewhere else to run: a microVM-isolated sandbox with Deny-default egress and no ambient identity, reached as an ordinary tool call so the agent framework's middleware still sees the call and classifies its result — only the work leaves the process.

This package is the backend only, with no sandbox kind of its own. maf-sandbox-bicep is the first kind that runs on it, written against maf-sandbox's protocol rather than against this backend.

Quickstart

pip install maf-sandbox-acas

from maf_sandbox_acas import AcasSandboxBackend, AcasSandboxConfig
from maf_sandbox import SandboxRouter

backend = AcasSandboxBackend(AcasSandboxConfig(endpoint="https://management.<region>.azuredevcompute.io", subscription_id="<sub-id>", resource_group="<rg>", sandbox_group="<group>", registry="<acr>.azurecr.io"))
router = SandboxRouter([backend])  # microVM isolation meets the router's default floor

samples/01_acas_bicep runs that pair end to end: the same two lines, plus the caller context and the workload tool they exist to serve, in a program that validates a Bicep file and disposes the sandbox afterwards.

azure-containerapps-sandbox — the data-plane SDK this backend calls — is a hard dependency (it is still a preview, 0.1.0bN, package; pin it in your own lockfile if you need reproducibility beyond the range this package declares). Authentication is DefaultAzureCredential; see Azure Identity's docs for how it resolves credentials in your environment.

Threat model

The micro-VM boundary. AcasSandboxBackend declares Isolation.MICROVM: execution happens in a hardware-isolated microVM, not a shared-kernel container, and that rung is maf-sandbox's router's default floor — a host that configures nothing already permits this backend (see that package's README). Everything below this line assumes that boundary holds; it is a property of the Azure Container Apps Sandboxes service, not of this package's code.

What identity is reachable. No ambient identity is placed inside the sandbox — the control-plane credential this package uses to create and manage sandboxes (DefaultAzureCredential) never travels into the guest. Code running inside a sandbox has no path back to the host's Azure identity, the host process's environment, or any other conversation's sandbox: dispose_scope deletes by service-side label, not by trusting the caller, and egress is Deny-default with a per-spec allowlist supplied by the kind, not by runtime configuration — a deployment that could widen a kind's egress after the fact could undo the containment its design rests on.

The backend

AcasSandboxBackend implements maf_sandbox.SandboxBackend:

Two image namespaces, and spec.image says which by whether it carries a tag. The service prebuilds images and keeps them Ready for every sandbox group — python-3.13, node-22, ubuntu and a dozen more — and a spec reaches them by naming one, with no registry and no tag, because the version is part of the name. Anything else is the repository:tag the rest of this package is written around: qualified by the configured registry and resolved against the disk images this deployment imported with scripts/import_disk_image.py.

codeact_sandbox_spec(image="python-3.13")           # the service already has it — import nothing
bicep_sandbox_spec(image="bicep-sandbox:0.46.1")    # yours, imported once, qualified by `registry`

The tag is what separates them, and it has to be: bicep-sandbox:0.46.1 names no registry either, so a rule that looked only for a registry would swallow every deployment configuring an imported image the way SandboxSpec documents. A bare name the service does not have is refused before anything is created, with the catalogue in the message — the likely way to arrive at one is a forgotten tag, and the fix is then visible where the error is. image_id still skips both lookups, as the field promises.

Microsoft's docs call these public images, glossed as "prebuilt images available to all sandbox groups", in the same paragraph that calls Docker Hub a public registry. This package says prebuilt to keep those apart; the SDK spells them list_public_disk_images() and begin_create_sandbox(disk=…).

tests/test_acas_e2e.py is the live suite, skipped unless ACAS_SANDBOX_ENDPOINT and MAF_SANDBOX_ACAS_E2E_IMAGE name a sandbox group and a guest image. It is what exercises the real data plane — the shared FILES_OUT conformance probes, the cap and confinement refusals, the read timeout that turns a fifo from a hang into a refusal, and teardown read back from the service rather than from this process's memory. It runs in verify-live.yml, not on a pull request, because every sandbox in it is billable. Most of it shares one sandbox; the prebuilt-image probes need a second, booted from python-3.13 (override with MAF_SANDBOX_ACAS_E2E_PREBUILT), because a name from the catalogue is the thing they exist to prove boots.

Capability.FILES_LIST as well as FILES_OUT, and this is the only backend that declares it. The service enumerates a directory natively, which is the test the protocol's split applies — name the backend that lacks it. A kind whose output names are unpredictable is refused on Docker and wslc and served here.

Capability.HOST_TOOLS, and what it claims is narrower than the others. It is the only member of the vocabulary with no backend method behind it — the transport is composed by the kind out of exec, write_file, stat_file and read_file, all of which the capabilities above already cover. What this backend adds by declaring it is one property: exec detaches. A process started by one call outlives it and is still observable from the next, because the sandbox is a microVM the group keeps between calls, and host_tool_calls_over_exec is built on exactly that — its launcher returns at once and the appearance of the exit-code file is the run's only witness. That is measured against the service, not asserted: TestWhetherThisBackendCouldServeHostTools in the live suite watches the exit marker be absent when the launcher's exec returns and appear afterwards. It is not a claim about the image — the shipped launcher wants sh, nohup, printf, mv, mkdir, rm and kill, and setsid where the image has it, and a kind wants whatever interpreter it names, none of which this backend chooses (#111).

An image whose guest is not root cannot serve FILES_OUT or HOST_TOOLS, and acquire says so. The two planes act as two principals: write_file lands files as 0:0 whatever the image's USER is, while exec runs as that USER and the SDK offers no selector to raise it. Every directory the file plane creates is therefore root-owned and 0755, and a guest program on a non-root image can create nothing inside one — not a declared output, and not the pid and exit markers the host-tool launcher writes, which is why that transport cannot even start there (#722). So acquire reads the guest's uid once per image, refuses a spec requiring either capability, warns about one requiring only EXEC — a command whose whole result is its stdout is unaffected — and serves an image whose uid it could not read, because refusing on an unreadable probe would take a working root image off a deployment. The wall is a directory the file plane made rather than the working directory as such — /tmp stays writable — and it bounds every shipped kind because each reaches the guest through a write_file that creates the call directory. It lands at acquire rather than at attach for the reason every image-shaped refusal does: nothing is running when the router matches a capability set.

Only regular files are read, and the refusal happens at stat time. This backend's read follows symlinks: a path linking to /etc/hostname returns that file's contents, so classifying after the bytes come back would be too late. The type comes from the data-plane payload's isSymlink and isDir flags, read raw — the SDK's typed FileInfo exposes neither, and a payload missing them is refused as AcasEntryPayloadIncomplete, never assumed regular (#136).

What the type check cannot prove. isDir and isSymlink establish that an entry is neither of those; they do not establish that it is a regular file, and mode is permission bits with the type stripped. A FIFO is reported identically to an empty regular file and is classified FILE — and reading one never returns, so read_timeout_seconds bounds it and a hang becomes a refusal rather than a held-open turn. The missing signal is filed as microsoft/azure-container-apps#1807.

Every path component is checked, not just the last one. A guest that points out at /etc gets a stat of out/hostname that says "regular file, 12 bytes" — the parent link is invisible there — so stat_file, read_file and list_dir stat every parent component from the filesystem root down, not from the working directory, whose own ancestors the guest can replace just as easily: with /maf-sandbox -> / unchecked, /maf-sandbox/work stats as a real directory and serves /. A link anywhere among those ancestors is refused as an escape; any other non-directory is an ordinary ENOTDIR. The check is maf_sandbox.paths.refuse_symlinked_ancestors, shared with every other backend serving FILES_OUT; what this package supplies is the unconfined, no-follow stat it runs on. Only the parents are refused — a link as the final component is still described as SYMLINK, which is how a caller learns it is one. One residual stays open and cannot be closed with this API: the stat and the read are separate calls, and the service has no no-follow read, so a guest that swaps a stat-ed file for a symlink in between is followed.

That dispose_scope detail is the one worth reading twice. A multi-replica host serves a conversation delete wherever it lands, so the replica that created a sandbox is usually not the one deleting it. A backend that consults only its own registry leaves billable sandboxes running, and the bug is invisible on a single-replica dev box. Sandboxes are labelled at create time so the service can answer the question instead.

Egress comes from the spec, not from configuration: default_action: Deny plus one Allow rule per host the kind declares. A deployment that could widen a kind's egress could undo the containment its design rests on.

Upgrading to 0.15

The four optional declarations moved into one BackendDeclarations. maf-sandbox 0.26 replaced capabilities, limits, egress_modes and os_families as backend attributes with one declarations object holding them as fields, and this backend follows it. A host that read them off the backend gets an AttributeError:

Nothing about what this backend declares changed — the values, and how they are derived from the config, are exactly as they were. maf-sandbox's own README carries the reasoning and what a backend author has to do.

Extracting this package

It imports nothing from its host application — only maf-sandbox and azure-* — so moving it to its own repository is a file move plus a dependency line. src/, tests/, scripts/ and pyproject.toml are already the future repo root.

TestOnlyDeclaredDependencies is what keeps that true: it scans this package's sources and fails on any import that is neither the standard library, this package itself, nor a distribution its own pyproject.toml declares. Nothing else would notice a stray one, because a workspace has every sibling already on the path — and an undeclared import is exactly what breaks a fresh pip install of the published wheel.

What stays behind is the host's adapter — a single module in the host application that maps the host's settings onto an AcasSandboxConfig and supplies the request context. Read it first if you want to know what integrating this package involves.

Provenance

Extracted from a production agent application, where a security review chose a microVM-isolated sandbox over running agent-authored code in the host process. Both halves of that conclusion are visible in this backend's design: the boundary it declares, and the fact that no credential of the host's ever travels inside it.

Maintained by SOKOLAI BV.