Skip to main content

Why RAI Planning?

A Different Kind of Risk

Traditional security planning focuses on how adversaries exploit systems. Injection attacks, broken access control, cryptographic failures: these are threats where an attacker deliberately subverts intended behavior. Security assessments excel at finding and mitigating these risks.

AI systems introduce a fundamentally different risk category. A model can produce biased outputs, make opaque decisions that affect people's lives, or amplify existing societal inequities, all while functioning exactly as designed. No attacker is required. The system's own behavior is the risk.

The Core Insight

Security planning asks: "How can this system be attacked?"

RAI planning asks: "How can this system cause harm, even when working correctly?"

Both questions matter. They require different analytical frameworks, different threat taxonomies, and different mitigation strategies. Running a security assessment alone leaves the AI-specific security model unexamined.

How Each Phase Addresses AI Risk

Phase 1: AI System Scoping

Establishes what the AI system does, who it affects, and where it operates. Identifies all AI and ML components, their data inputs, and the decisions they influence. Without clear boundaries, risk assessment cannot be comprehensive.

Phase 2: Sensitive Uses Assessment

Screens the system against Microsoft's sensitive uses categories, including applications in healthcare, criminal justice, employment, and education where AI decisions carry elevated consequences. Catches restricted use scenarios that require organizational escalation before development continues.

Phase 3: RAI Standards Mapping

Maps each AI component against the six Microsoft RAI principles and NIST AI RMF subcategories. Identifies which principles apply to which components and what regulatory obligations exist. This mapping becomes the evaluation framework for Phases 4 and 5.

Phase 4: RAI Security Model Analysis

Applies AI-specific threat analysis across seven categories: data poisoning, model evasion, prompt injection, output manipulation, bias amplification, privacy leakage, and misuse escalation. These categories reflect how AI systems fail, not how traditional applications are exploited.

Phase 5: RAI Impact Assessment

Evaluates whether adequate controls exist for each identified threat. Documents evidence of mitigations already in place and identifies gaps. Analyzes tradeoffs where RAI principles compete, such as transparency versus privacy or fairness across different demographic groups.

Phase 6: Review and Handoff

Produces a quantified scorecard across five dimensions and converts gaps into actionable backlog items. The score determines whether the system proceeds, requires conditions, or needs remediation.

Quality Comparison

DimensionAd-hoc assessmentRAI Planner
Threat coverageVaries by assessor expertiseSeven AI-specific threat categories applied systematically
Standards traceabilityOften informal or missingEach finding mapped to RAI principles and NIST AI RMF subcategories
Sensitive uses screeningFrequently overlookedMandatory Phase 2 gate with restricted uses escalation
ReproducibilityDepends on individual documentationStructured state, artifacts, and scoring produce consistent results
Backlog integrationManual translation to work itemsAutomated generation with autonomy tiers and dual-platform support
Security plan continuitySeparate process, no shared contextfrom-security-plan mode inherits AI components and threat sequences

Learning Curve

You do not need RAI expertise to start. The agent guides the assessment conversationally, asking focused questions and explaining concepts as they arise.

StepActivity
First sessionRun in capture mode on a project you know well. Answer the questions naturally and review the generated artifacts.
Second sessionTry from-security-plan mode after completing a security plan. Notice how AI component data carries forward automatically.
Third sessionReview the scorecard and backlog output. Use the generated work items to drive actual mitigations.
OngoingThe structured artifacts serve as living documentation. Return to update assessments as the AI system evolves.

Choosing Your Approach

Factorcapturefrom-prdfrom-security-plan
Starting contextNone; full interview from scratchPRD or BRD artifacts in .copilot-tracking/Completed security plan with AI components
AI component discoveryManual during Phase 1 questionsExtracted from product documentationPre-populated from security plan state
Threat ID continuityStarts at RAI-T-{CATEGORY}-001Starts at RAI-T-{CATEGORY}-001Continues from security plan threat count
Time to Phase 2Longest: full scoping interviewMedium: confirm and refine extracted scopeShortest: verify pre-populated data
Best forExploratory assessments, standalone AI projectsProjects with existing product documentationThe recommended workflow after security planning

See Entry Modes for step-by-step instructions on starting each mode.

🤖 Crafted with precision by ✨Copilot following brilliant human instruction, then carefully refined by our team of discerning human reviewers.