Skip to content

Security and support scope

pVisor aims for bounded, reversible, and checkable execution: cap the blast radius of mistakes, let you discard file changes before apply, and record truthfully which controls actually took effect.

At a glance

Protects Does not protect
Staged workspace: project files stay unchanged until apply, and conflicts refuse to overwrite your edits External effects that already happened: remote API calls, database writes, sent messages
Sensitive paths: the --safe preset rejects .ssh, .gnupg, and private-key files inside the view Secrets outside the view, explicitly shared, or renamed; keys in source and Git history
Mandatory network boundaries: VM auto, host --overlaynet-deny-all, container offline mode Clients that ignore the proxy or open raw sockets under a cooperative proxy
Honest evidence: separates the request, the plan, the installed controls, and the observation Kernel bugs, side channels, misuse of explicitly granted credentials

Each Run's actual boundary comes from its own capability evidence, not from configuration, preset names, or executor names.

pVisor is neither cryptographic remote attestation nor hostile multi-tenant isolation.

Platforms, interfaces, and capability status

When choosing an execution environment, check prerequisites in Platforms and executor support, then verify that Executor boundaries cover the required file, network, and process controls. Native Jobs, standalone containers, and the daemon expose different interfaces. See the integration matrix for native Job, Runtime, daemon, and shim capabilities, and the CLI reference for standalone container commands.

Check requested policy, planned controls, and installed controls separately using Capabilities, evidence, and guarantees. Experimental capabilities such as automatic wait management have their own implementation stages; experimental results do not establish support across all platforms or interfaces.

Review the security of one task

List the files, credentials, networks, and external services available to the task. Establish the trusted side using the Threat model, and configure controls using Hardening. After a native Job runs, use pvisor status --review to inspect installed controls, warnings, and file changes. Consult Known limitations before accepting results. Missing observations remain unknown; cooperative controls do not establish mandatory isolation.

When you discover a boundary problem

If a task crosses a boundary its own evidence marks Enforced, or a record misreports controls, privately submit the version, executor, command, and minimal reproduction using Vulnerability disclosure. Do not disclose vulnerability details in a public issue.

No independent security audit report is currently registered publicly. Internal regressions and Run Bundles help check implementation and execution results but do not replace an independent assessment. See Third-party audits for assessment scope and report requirements.