The decision
is human
AI can now write the security review. The reasoning that used to take days is becoming abundant. The step that has not moved is the decision.
Read the essayAI Security Review
Security clearance for every AI system
Drel reviews your AI system, clears it to ship, and keeps that clearance backed by evidence as your code changes. Ship AI you can prove.
Built on established AI security frameworks
OWASP LLMOWASP AgenticNIST AI RMFISO 42001EU AI ActMITRE ATLASMAESTRO
Works with the stack you already use.
Drel helps security and product teams review AI, RAG and agentic systems built with the model providers, cloud platforms and engineering tools they already use.
The problem
AI teams ship faster than security can assess.
AppSec and security architecture teams are not staffed to manually threat-model every AI, RAG, or agentic system. Traditional threat modeling tools were not designed for LLM trust boundaries, retrieval authorization, or agentic tool use. Assessments pile up. Systems go live without proper security sign-off.
The solution
Turn AI architecture into a defensible security decision.
Describe your AI system. Review its architecture, controls, and evidence. Drel turns the review into a structured decision with clear requirements, ownership, and sign-off, defensible in front of an AI Committee, regulator, or board.
Purpose-built
Security review built for AI systems.
Review RAG pipelines, agentic workflows, and LLM-powered applications through a security model built around AI architecture, authority, data flows, tools, and human approval.
Designed around AI-specific architecture and risk, rather than forcing AI systems into generic application-security workflows.
What you get
A clearance decision backed by evidence.
Every output is specific to your AI system: named blockers before production, required controls with owners and deadlines, evidence gaps that must close before go-live, and a sign-off chain that creates an audit trail.
Clearance decision: proceed, conditional, restricted pilot, hold, or decline
Evidence on every claim: explicit, inferred, assumed, unknown, missing, or verified
Production blockers with required controls, owners, and deadlines
Re-assessment triggers that fire when AI systems change
Multi-stakeholder sign-off chain: CISO, AI Governance, DPO, Business Owner
Audit-ready record: versioned, timestamped, framework-mapped
Evidence
Find the evidence in the code.
Link the GitHub repository behind the system. Drel looks for the code or configuration that shows each control is in place, reads every file at one commit, and quotes the exact line with a link to it. Your reviewer decides whether it counts as evidence.
Drel reads a bounded set of files and stores nothing from the repository. A finding is a quoted line for a reviewer to judge: Drel never marks a control verified on its own.
- 01
Describe
The system, its data, and what it is allowed to do.
- 02
Find
The code that shows each control is in place, quoted from your repository at one commit.
- 03
Decide
Your reviewers judge what was found, then approve, restrict or hold, with the evidence attached.
After sign-off
From a change to a new decision.
When the system changes, record it in Drel. Drel reopens only the parts the change affects, gives each one an owner, and shows what would close it. The clearance you already signed stays exactly as it was signed, and the new version gets its own decision.
Changed
You record a change to the system. Drel reopens only the parts it affects.
Owned
Each reopened item has an owner and a due date.
Closed
The new version is reviewed and sealed, with its evidence attached.
Signed clearance · stays exactly as signed
The signed clearance stays exactly as signed while the correction runs.
A signed clearance stays a record of the basis it was signed on. When that basis changes, Drel shows you the difference. It does not rewrite history, and it does not renew or revoke a clearance on its own.
Behind the review
Our practice includes individual work in OWASP GenAI Security, CSA AI Safety, and AIUC-1.
About Drel →121
Attack paths modeled
Prompt injection, tool-chain escalation, approval bypass, confused deputy. Pre-modelled for every AI system type.
12
Frameworks mapped
OWASP LLM Top 10, NIST AI RMF, EU AI Act, ISO 42001, and more. Linked to every decision.
5
Clearance states
Proceed, conditional, restricted pilot, hold, or decline. Every review ends with a decision, not a report.
<10 min
To a clearance decision
From system description to a decision your AI Committee can sign, without routing every system through a manual review queue.
Differentiation
Security reviews that end with a decision.
Drel turns AI system reviews into defensible clearance decisions, linking blockers, evidence, ownership, sign-offs, and re-review triggers into one audit-ready record.
Clearance
Clear what can ship
Turn review evidence into a clearance decision: proceed, conditional, restricted pilot, hold, or decline.
Blockers
Expose what blocks release
Separate advisory findings from production blockers, missing gates, and unresolved assumptions.
Evidence
Grade the evidence
Track whether each claim is explicit, inferred, assumed, missing, or verified.
Audit Record
Leave the audit trail
Capture rationale, owners, sign-offs, versions, and re-review triggers in one defensible record.
Coverage
Built for every AI system type.
RAG assistants, tool-using agents, customer-facing AI features: each architecture pattern carries distinct trust boundaries, retrieval risks, and control requirements. Drel maps all of them to the blockers, evidence states, and clearance decision they require.
Assistants & RAG
- Internal RAG assistant
- Customer-facing chatbot
- LLM gateway
Agents & automation
- Agent with tools
- Agentic automation
- Multi-agent workflow
Product & vendor AI
- B2B SaaS AI feature
- Vendor AI assessment
- Embedded AI capability
Reviewed through one clearance model: blockers, evidence states, control ownership, sign-offs, and re-review triggers.
System types
Every architecture, its own threat model.
Each system type has its own threat model, risk patterns, and control library. Built from the specific trust boundaries of that architecture.
Internal RAG Assistant
Retrieval-Augmented Generation over enterprise knowledge
Employee-facing assistants over SharePoint, Confluence, and ServiceNow introduce unique trust boundaries: retrieval authorization, prompt injection via documents, and identity propagation across the retrieval chain.
OWASP LLM01Indirect prompt injection via indexed documents
MITRE AML.T0054Retrieval ACL bypass: chunks returned without permission check
MAESTRO L3Identity not propagated from user session to retrieval layer
Agent with Tools
Agentic systems with write access and tool execution
LLM agents with API access to GitHub, Slack, Jira, and PagerDuty require explicit policy gates, approval boundaries, and action authorization. Without them, a single injected instruction can trigger cascading write actions.
OWASP LLM08Excessive agency: write actions without human approval gate
MITRE AML.T0051Tool call injection via adversarial user input
MAESTRO L5Privilege escalation through chained tool invocations
B2B SaaS AI Feature
Customer-facing AI in multi-tenant SaaS products
AI features embedded in B2B SaaS products must enforce strict tenant isolation, prevent cross-tenant data leakage, and produce go-live evidence for enterprise security questionnaires.
OWASP LLM06Tenant data isolation failure: context bleeds across sessions
MITRE AML.T0048PII exfiltration via LLM output in shared inference
NIST AI RMFMissing go-live evidence for enterprise security assessment
Who uses Drel
Built for the people responsible for AI security.
Whether you're a security architect running a solo review or an AI governance lead producing evidence for a committee, Drel gives you the structured clearance record you need.
Security Architects
Clear AI systems on your own.
Produce a defensible clearance decision with the controls and evidence behind it. No committee required.
Explore for security architects →AppSec EngineersAI review, inside your SDLC.
Map AI-specific risks and controls into your existing AppSec workflow and control library.
Explore for AppSec →AI Governance & DPOsEvidence your committee can sign.
Give committees the clearance record, control gaps, and evidence they need to sign, for ISO 42001 and the EU AI Act.
Explore for governance →Topic guides
AI security reviewThe end-to-end clearance workflow
RAG security assessmentIngestion, vector store, retrieval, prompt assembly
Agentic AI securityTool surfaces, delegation chains, blast radius
MCP security reviewModel Context Protocol deployments
Vendor AI assessmentThird-party AI before procurement
OWASP LLM Top 10Controls mapped to all 10 categories
OWASP Agentic Top 10Controls for agentic AI threat categories
AI Governance FrameworkISO 42001 clauses to evidence and controls
EU AI Act classifierUpdated for the 2026 AI Omnibus
EU AI Act change mapBefore/after dates and operational scope
Framework coverage
The standards your team already uses.
Every threat, control, and remediation item maps directly, so the output lands in your assessment without translation.
OWASP LLM Top 10Security risks in LLM-powered applications
OWASP Agentic Top 10The ten most critical risks in AI agents
OWASP AIVSSHow agentic capabilities amplify risk
MITRE ATLASAdversarial tactics against AI systems
MAESTROLayer-by-layer threat modelling for agentic AI
IBM AI Risk AtlasTaxonomy of ML, AI and agent risks
Cisco AI SecurityAttacker objectives and subtechniques
NIST AI RMFGovern · Map · Measure · Manage
ISO/IEC 42001AI management system standard
EU AI ActObligations for high-risk AI systems
AIUC-1Security, safety and reliability for AI agents
CSA AICMSecurity controls for AI/ML systems
From the essays
The decision
FAQ
Common questions before the first review.
- What is an AI security review?
- A structured, evidence-backed evaluation of an AI system before it reaches production: a threat model specific to the system's architecture, the controls required to address each threat, the evidence gaps still open, and a clearance decision. Not a slide deck, and not a vendor's self-attestation.
- How is this different from a penetration test?
- A penetration test probes a system that already exists, in a running environment, for exploitable vulnerabilities. An AI security review happens at design time, before or alongside build, and produces a control plan and a clearance decision rather than a findings list. The two are complementary, not substitutes: a design-time review misses what only shows up at runtime, and a pentest misses training-data and disposition-level risks that never appear as a runtime exploit.
- What decision does a review produce?
- One of five states: proceed (unconditional), conditional (proceed with named conditions), restricted pilot only (limited deployment with a re-review trigger), hold (more work required before any deployment decision), or decline. A binary pass/fail model doesn't capture how AI systems actually clear review in practice.
- Does Drel certify AI systems?
- No. Drel produces a structured, evidence-backed assessment: the record of what was reviewed, what was accepted, and under what conditions. Certification is a separate, accreditation-based process performed by audit bodies, not by review platforms.
- What frameworks does Drel map to?
- OWASP LLM Top 10, OWASP Agentic Top 10, ISO 42001, EU AI Act (Article 9 risk management and Annex III classification), NIST AI RMF, and MITRE ATLAS, mapped to what the specific assessed system actually does rather than applied as a fixed checklist. Sector-specific frameworks (HIPAA, PCI DSS, GDPR, SOX) apply when the system touches that kind of data or activity.
- Does this replace our AI Committee?
- No. Drel produces the evidence pack and disposition memo the AI Committee reviews and signs off on: the input to the Committee's decision, not a replacement for it. Named accountability for accepted residual risk stays with the Committee, not with the tool that produced the evidence.
Ready to secure your next AI system?
See how Drel produces a defensible clearance decision for your AI Committee: threats, controls, evidence, and a go/no-go recommendation.