SovereignRoot blog
AI agent boundaries, honestly explained
Practical, no-hype writing on signed policy files, prompt injection, and what actually enforces a boundary.
Agent delegation and authority amplification
Child agents can silently amplify authority. Attenuation-only delegation and a root ceiling keep scope creep in check.
Signing JSON in the browser with Web Crypto
ECDSA P-256, RFC 8785 canonicalization, JWK thumbprints, and encrypted key export.
AI agent authorization: the grant layer and the ceiling layer
OAuth, delegation receipts and scope tokens grant authority. A signed root policy caps it.
What a signed AI policy file can and can't do
A signed sovereignty.json is tamper-evident and portable. It is not, by itself, a security boundary.
How to define AI agent boundaries that survive model changes
Deny-overrides, require-approval, and attenuation-only semantics — a practical guide.
Prompt injection, policy files, and the enforcement gap
Putting a policy in a prompt is not the same as enforcing it. Here's the gap.