Pre-production AI clearance
Before an AI system handles real users or production data, you need a structured security gate. Drel produces the clearance decision, control verification record, and evidence pack that makes that gate defensible.
Why a clearance gate matters
AI systems fail in ways that traditional software doesn't. Prompt injection, retrieval poisoning, tool misuse, data leakage through memory — these are not bugs you catch in a standard code review. They require a different kind of pre-production review.
A pre-production clearance answers one question: is this system ready to handle real users and real data, and what conditions apply? Without a structured answer, teams ship on confidence rather than evidence. The difference matters when something goes wrong.
The two gates
Most production AI systems require at least two clearance gates:
- Pre-pilot clearance. Before the system is exposed to real users, even in a limited pilot. The threat model is complete, critical controls are in place, and conditions for the pilot are named.
- Pre-production clearance. Before the system handles production traffic and production data. All control gaps from the pilot are resolved or explicitly accepted as residual risk with an owner.
Drel supports both gates. The same assessment workspace carries the system from initial review through to the production clearance record.
What Drel produces
For a pre-production clearance, Drel produces:
- A system-specific threat model — not a generic checklist, but threats identified for the architecture, tools, and data flows of this system.
- A required controls list — every control that must be in place before pilot, and every control required before production.
- A control gap analysis — controls not yet implemented or evidenced, with explicit lifecycle gate assignments.
- A clearance decision — one of five states (Proceed, Conditional, Restricted Pilot, Hold, Decline) with named conditions and re-assessment triggers.
- An evidence pack — exportable PDF with threat model, control plan, gap list, disposition memo, and sign-off record.
What triggers a re-assessment
A clearance is valid for the system as assessed. The following changes trigger a re-assessment:
- Model or model version change
- New tool or MCP server added to the agent
- Change in training data or fine-tuning
- Scope expansion — new data types, new users, new use case
- Security incident involving the system or its components
- Scheduled periodic review (typically annual for production systems)
Each re-assessment trigger should be named in the original clearance decision, with an owner. Drel records these triggers in the clearance record so they don't get lost.
How long a review takes
A structured AI security review in Drel takes 2–4 hours for an experienced security architect working from a complete system description. The main source of delay is incomplete intake — if the team can't answer questions about tool permissions, data flows, or hosting model, the review will stall.
Drel's system intake guide lists the information required for a complete review before you start. Running through it with the building team before the security review is the most effective way to reduce review time.