Who?
Which agent or identity requested the capability?
AI SECURITY / AUTHORITY CONTROL
Watchdog
Prompt injection should not equal authority.
Watchdog is being developed as a policy enforcement boundary between an AI agent's reasoning and the tools capable of causing real-world side effects.
ARCHITECTURE
Intelligence requests. Policy decides.
UNTRUSTED INPUT
→
AGENT
→
CAPABILITY REQUEST
→
WATCHDOG
→
TOOL BROKER
The intended design prevents agents from directly reaching privileged tools. Requests pass through an explicit policy decision point first.
POLICY QUESTIONS
Every action has context.
What?
Which capability is being requested?
Where?
Which resource or trust domain is targeted?
Why?
What input and provenance caused the request?
DECISIONS
ALLOW
Request satisfies the active policy.
DENY
Capability or resource is not permitted.
HUMAN AUTHORIZATION
High-consequence actions can require explicit human presence.
Try attacking the model.
The public demonstration uses simulated capabilities and fabricated data.
Launch Attack Demo