Everest Security
One governed edge, optional surrounding controls.
Reference architecture for identity-aware enterprise AI security investigation with optional Aembit, CrowdStrike AIDR, Falcon MCP and live log evidence.
Layers
| Layer | Role |
|---|---|
| AI client / BrainStorm | Conversation, reasoning, local or hosted model, scheduled agent loop. |
| Aembit (optional) | Human + agent access policy and credential brokerage. |
| Steward | Authentication, agent attribution, backend exposure, policy, audit and local STDIO runtime. |
| AIDR (optional) | MCP content inspection and allow/block/redaction decisions. |
| Falcon MCP | Security-platform context under Falcon API scopes. |
| Logs MCP | Live distributed application/OS evidence, including security evidence records. |
Interactive and unattended use
Interactive users can reach supported AD resources under their own authority. SecurityWatch runs under a dedicated automation identity and can be restricted to a separate security-profile backend.
Design principle
Ask at every hop: which identity or credential does the target resource actually authorise?