Partly real, but that wording smells like model overconfidence.
OpenClaw does have real sandbox/workspace boundaries. In v2026.5.x, inbound attachments are staged into media/inbound/* in the active workspace for the agent/sandbox — not necessarily readable as the host path ~/.openclaw/media/inbound/... from inside the agent.
So if the agent tries to read .openclaw/media/inbound/... or ~/.openclaw/media/inbound/..., it may fail because it is outside the sandbox/workspace it can see. That is expected. But “hardcoded security boundaries” is a fuzzy explanation; the practical issue is usually path/sandbox config, not Gemini discovering some special secret restriction.
Try this first:
- Ask it to use the path shown in the inbound attachment, or relative path like
media/inbound/<filename> — not ~/.openclaw/media/inbound/....
- If it’s an image/PDF/audio, ask it to use the dedicated
image / pdf / attachment-aware tool rather than shell-reading the host media folder.
- Check your config for sandbox settings. With sandboxing enabled,
agents.defaults.sandbox.workspaceAccess controls what workspace the agent can see. Docs: https://docs.openclaw.ai/gateway/sandboxing
- If this is still failing, run:
openclaw status --all
openclaw logs --follow
and look for the exact file path/error.