Blog

Notes on building an AI coding agent a security review can actually approve.

• Enforced vs advisory: what "sandboxed" really means in an AI coding agent — Most agents describe safety features they cannot enforce. Where we draw the line, and why we publish the gaps.