Handoff Pipeline

Security Planner to RAI Planner Pipeline

The Security Planner and RAI Planner form a connected assessment pipeline. When the Security Planner detects AI or ML components during Phase 1, it captures component details and enables RAI dispatch. At Phase 6, the Security Planner recommends starting the RAI Planner in from-security-plan mode.

flowchart LR
  subgraph SP ["Security Planner"]
    SP1["Phase 1\nScoping"]
    SP6["Phase 6\nHandoff"]
    SP1 -->|"detects AI components"| SP6
  end

  subgraph Handoff ["State Transfer"]
    ST["state.json\naiComponents\nraiEnabled\nraiScope\nraiTier\nthreatCount"]
  end

  subgraph RP ["RAI Planner"]
    RP1["Phase 1\nAI System\nScoping"]
    RP6["Phase 6\nReview &<br/>Handoff"]
    RP1 --> RP6
  end
  SP6 -->|"recommends entry"| ST
  ST -->|"pre-populates"| RP1

What the RAI Planner Receives

When entering via from-security-plan mode, the RAI Planner reads the security plan's state.json and inherits:

Data	Source field	How it is used
AI component inventory	aiComponents array	Pre-populates Phase 1 AI element catalog
RAI assessment scope	raiScope	Sets initial assessment boundaries
RAI depth tier	raiTier	Determines assessment depth (Basic, Standard, or Comprehensive)
Threat count	threat catalog size	Starting sequence for T-RAI-{NNN} IDs
Security plan reference	state.json path	Stored in securityPlanRef for cross-referencing

NOTE

The RAI Planner reads security plan artifacts as read-only. It never modifies files under .copilot-tracking/security-plans/.

Review Summary Generation

Phase 6 produces a review summary covering observations across six dimensions. The review summary presents maturity indicators rather than numeric scores, supporting handoff decisions and backlog prioritization.

flowchart TD
  P1["Phase 1\nAI System Scoping"] --> D1["Standards\nAlignment"]
  P2["Phase 2\nRisk Classification"] --> D6["Risk\nClassification"]
  P3["Phase 3\nStandards Mapping"] --> D1
  P4["Phase 4\nSecurity Model"] --> D2["Threat\nCompleteness"]
  P5["Phase 5\nImpact Assessment"] --> D3["Control\nEffectiveness"]
  P5 --> D4["Evidence\nQuality"]
  P5 --> D5["Tradeoff\nResolution"]

  D1 --> RS["Review Summary"]
  D2 --> RS
  D3 --> RS
  D4 --> RS
  D5 --> RS
  D6 --> RS

Review Dimensions

Dimension	What it covers
Standards Alignment	How well AI components map to trustworthiness characteristics and regulatory frameworks
Threat Completeness	Completeness and accuracy of AI-specific threat identification
Control Effectiveness	Coverage and effectiveness of controls for identified threats
Evidence Quality	Quality and availability of evidence supporting control effectiveness
Tradeoff Resolution	Clarity of characteristic tradeoff analysis and resolution rationale
Risk Classification	Whether risk indicators were evaluated with documented mitigations

Backlog Generation

Gaps identified during Phases 3 through 5 are converted to work items using the same dual-platform format as the Security Planner.

Dual-Platform Support

Platform	ID format	Formatting	Target system
ADO	WI-RAI-{NNN}	HTML <div> wrapper	Azure DevOps work items
GitHub	{{RAI-TEMP-N}}	Markdown with YAML frontmatter	GitHub issues

Autonomy Tiers

Each generated work item receives an autonomy tier based on the severity and complexity of the finding.

Tier	Human involvement	When assigned
Full	Agent creates and submits without confirmation	Low-severity findings with clear remediation
Partial	Agent creates items; user confirms before submission	Default tier for most findings
Manual	Agent recommends; user creates items	High-severity findings or cross-team coordination required

Content Sanitization

All generated backlog content is sanitized before handoff:

1. Replace .copilot-tracking/ paths with descriptive text
2. Replace full file system paths with relative references
3. Remove state JSON content or references
4. Remove internal tracking IDs that are not work item IDs
5. Preserve standards references in all cases

Pipeline Artifacts

Artifact	Path	Generated during
System definition pack	.copilot-tracking/rai-plans/{slug}/system-definition-pack.md	Phase 1
Stakeholder impact map	.copilot-tracking/rai-plans/{slug}/stakeholder-impact-map.md	Phase 1
Risk classification screening summary	.copilot-tracking/rai-plans/{slug}/system-definition-pack.md (appended)	Phase 2
RAI standards mapping	.copilot-tracking/rai-plans/{slug}/rai-standards-mapping.md	Phase 3
RAI security model addendum	.copilot-tracking/rai-plans/{slug}/rai-security-model-addendum.md	Phase 4
Control surface catalog	.copilot-tracking/rai-plans/{slug}/control-surface-catalog.md	Phase 5
Evidence register	.copilot-tracking/rai-plans/{slug}/evidence-register.md	Phase 5
RAI tradeoffs	.copilot-tracking/rai-plans/{slug}/rai-tradeoffs.md	Phase 5
RAI review summary	.copilot-tracking/rai-plans/{slug}/rai-review-summary.md	Phase 6
Handoff summary	.copilot-tracking/rai-plans/{slug}/rai-backlog-handoff-summary.md	Phase 6

Artifact Attribution and Review

Persisted RAI artifacts include transparency footers that communicate AI involvement and establish expectations for human review. Footer composition varies based on whether an artifact is consumed by subsequent agent phases or delivered to human reviewers for validation and decision-making.

AI-Content Note

Every Phase 5 and Phase 6 artifact includes a transparency note at the end of the file:

Note - The author created this content with assistance from AI. All outputs should be reviewed and validated by a qualified human reviewer before use.

Human Review Checkbox

Artifacts delivered to human reviewers include a review validation checkbox beneath the AI-content note:

• Reviewed and validated by a qualified human reviewer

Reviewers check this box upon completing their assessment to signal that the content has been validated.

Full Disclaimer

The Handoff Summary and Compact Handoff Summary append the full RAI Planner disclaimer after the review checkbox. This disclaimer establishes that the agent is an assistive tool only and that all outputs require independent review by appropriate legal and compliance reviewers before use.

Footer Classification

The following table shows which footer components appear on each Phase 5 and Phase 6 artifact. Agentic artifacts are consumed by subsequent agent phases, while human-facing artifacts are delivered to reviewers.

Artifact	Phase	Category	AI-content note	Review checkbox	Full disclaimer
Control surface catalog	5	Agentic	✓
Evidence register	5	Agentic	✓
RAI tradeoffs	5	Human-facing	✓	✓
RAI review summary	6	Human-facing	✓	✓
ADO work items	6	Human-facing	✓	✓
GitHub issues	6	Human-facing	✓	✓
Transparency note outline	6	Human-facing	✓	✓
Monitoring summary	6	Human-facing	✓	✓
Handoff summary	6	Human-facing	✓	✓	✓
Compact handoff summary	6	Human-facing	✓	✓	✓

NOTE

Transparency note outline and monitoring summary are optional artifacts generated only when the user opts in during Phase 6.

End-to-end assessment flow

1. Security Planner completes Phase 6 with raiEnabled: true and AI component data in state
2. User starts RAI Planner with from-security-plan prompt, providing the security plan project slug
3. RAI Planner reads security plan state and pre-populates Phase 1 with AI components and threat count
4. Phases 1-5 proceed with focused assessment of AI-specific risks, building on the security plan's foundation
5. Phase 6 produces the review summary with observations across six dimensions
6. Backlog items are generated for identified gaps using the user's preferred platform format
7. Assessment artifacts persist under .copilot-tracking/rai-plans/{project-slug}/ for future reference and updates

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