Glossary

Security Approval

The decision the AI Committee signs on whether an AI system can reach production and under which controls: Approved, Conditional, Restricted pilot, On hold or Declined, backed by an evidence pack.

A security approval is what the AI Committee signs before an AI system reaches production. The AppSec lead or security architect runs the review; the AI Committee (Security Architecture, AI Governance, DPO, Business Owner and CISO delegate) signs the approval. It names what controls are in place, what evidence supports them, and what re-assessment triggers apply. It is not certification, which requires an accredited third party.

The decision is one of five states: Approved, Conditional, Restricted pilot, On hold or Declined. Approved means the decision is to proceed and the committee has signed. Most production-grade AI systems receive a Conditional approval on first review.

Drel produces the approval in five steps. Model: rebuild the AI system from its repository and design docs. Threat model: map every attack path to OWASP, MITRE ATLAS, the EU AI Act and more, with the controls each path requires. Attack: the attack paths through the system, mapped from its design and code: how an attacker gets from an entry point to a target, and the controls that block each path. Drel does not run attacks. Fix: Drel proposes the change in the Jira ticket and checks the merged pull request; it never writes to the repository. Monitor: the automatic checks you turn on are re-checked when the branch you link changes, with a gate in CI and the approval available to agents over MCP.

Approval is per-system. The approval for a procurement agent does not transfer to a customer-facing chatbot. A change that adds a model, provider or tool flags the system for re-review, and Drel re-analyzes only the attack paths and controls the change reaches.

Approval is dated and scoped. An undated approval lasts forever in theory and rarely in practice. An approval with no scope statement cannot be defended in an audit. A useful approval names what was reviewed, when, by whom, and under what conditions the decision applies.

Approval is not compliance. It supports compliance evidence: ISO 42001 clauses 8.2 and 8.3 (risk assessment and treatment records), EU AI Act Article 9 (risk management evidence) and GDPR Article 35 (DPIA where applicable). It does not certify any of those.