Why agent sandboxes fail: untested defaults leak host access
The article argues that many agent sandboxes only provide isolation where capabilities are actually removed, and that stated boundaries often fail in practice. It welcomes Docker's push into disposable agent sandboxes but stresses that isolation is a claim about defaults that must be verified. The author proposes a practical test: list what the agent cannot access on the host, then attempt to break out. Key checks include whether the Docker socket, home directory, read-only mounts, network egress, cloud metadata access, resource limits, user privileges, and persistence paths are truly restricted.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in