Epic: Agent Security & Capabilities

[tobiu]

Define and implement the security model for Agent-initiated browser actions.

Scope:

1. Capability Taxonomy: Define granular permissions (e.g., component:read, component:write, code:load).
2. Policy Enforcement Point (PEP): Implement middleware in Neo.ai.server.WebSocket to validate RPC calls against the Agent's capability token.
3. Sandboxing: Ensure Agents cannot execute arbitrary JavaScript (e.g., eval) in the browser context unless explicitly authorized.
4. Audit Logging: Record all Agent-initiated actions for security review.
5. Default Deny: All capabilities require explicit grant.
6. Emergency Kill Switch: Ability to revoke agent access immediately.

Reference: .github/AGENT_ARCHITECTURE.md