Alpha software. StarClaw 2.x is an alpha and is not recommended for customers, production use or anyone who needs a guarantee.
StarClaw alpha

Security: what is and is not protected

StarClaw does not prevent prompt injection. It assumes the model will sometimes be steered by hostile content. It limits what that can cost you: risky actions need a human, who is shown the exact action and where the untrusted content came from, and everything is provable afterwards.

This is an alpha. It has had no external review or fuzzing. Do not treat it as secure, production-ready or consumer-ready.

This page restates the "What is and is not protected" and "Known limits" sections of the README, and parts of SECURITY.md. If this page and those files ever disagree, the files in the repository win.

What is and is not protected

ThreatResult
Generated code spawns processes, loads addons/workers/WASI, opens sockets, reads or writes outside the workspace, signals other processesBlocked and probe-verified on Node 25.9.0 / Windows, in the process layer and (when available) again by the container layer. Details: starclaw doctor.
Generated code tries to reach the host (files, network, docker socket, environment, other processes)Container layer: no mounts, no network, no capabilities, non-root, read-only root, memory and process caps. Probe-verified from the host side on Docker Desktop (WSL2).
Injected text makes the agent exfiltrate data, run commands or spendNot prevented. Requires human approval once the session is tainted; the human sees the provenance.
You approve something harmfulNothing stops you. Approved host commands run with your full privileges.
Someone edits the audit log or memory without the signing keyDetected.
Someone who can read .starclaw/audit.keyCan forge both. The key sits next to the data.
Malicious code that finds a Node permission-model bypassWith the container layer: confined by the container instead. Without it: not covered (Node documents the model as a "seat belt").
Malicious code that finds a Linux kernel or container-runtime escapeNot covered. Containers share the kernel (on Docker Desktop: the WSL2 VM's). For hostile multi-tenant use, put gVisor, Kata or Firecracker underneath.

Known limits

Threat model in short

Considered: hostile content that tries to steer the model; a misbehaving or compromised model that emits arbitrary code; a local attacker who can write to the state directory but does not have the signing key.

Not considered: an attacker who can read the signing key or run code as your user outside StarClaw; a malicious maintainer or dependency; physical access; side channels.

Residual risks, plainly

  1. Prompt injection works against the model. StarClaw only gates what the model can do. A human who approves is the last line of defence.
  2. The process layer is a seat belt. A container or kernel escape is not covered.
  3. Approved host commands are unsandboxed.
  4. Key and state files are only as safe as your user account.
  5. Only one OS and Node combination was tested. No external review or fuzzing was done.

Reporting a problem

SECURITY.md asks that vulnerabilities be reported privately to the maintainer, and that exploits not be filed as public issues. Report them to security@starclaw.dev.