NEW

Start with the pressure: sales, launch, abuse, agents, data, or guardrails

Use the engagement router
aisecurityllc
Sign inStart AI Security Assessment
aisecurityllc
Start AI Security Assessment
← Deliverablesdeliverable
deliverable
public-sample

AI Red-Team Findings Register

A technical companion to the executive summary with structured findings, severity, reproduction summary, affected boundary, evidence, remediation, and validation status.

14-28 pages
Scope AI Red Team & Adversarial TestingScope AI Product Security AssessmentScope Agentic Workflow Security & Hardening
Client deliverable
public-sample
14-28 pages
System
Northstar Support Cloud / Customer Support Copilot
Environment
staging with production-like synthetic tenant data

# AI Red-Team Findings Register

Sample Deliverable

Executive Summary

This findings register is the technical companion to the AI red-team executive summary. It captures each validated finding, affected boundary, safe reproduction summary, evidence, business impact, remediation, validation criteria, owner, and release decision.

The public sample avoids payload-level exploit detail while preserving the structure needed for engineering remediation.

Heads up

Public sample notice

This is a shortened, synthetic excerpt prepared as a public sample. A client version would include system-specific evidence, implementation references, architecture screenshots, control test results, owner sign-offs, and full supporting documentation. This sample uses Northstar Support Cloud / Customer Support Copilot as the synthetic reference system. This sample is not legal advice, not a compliance certification, not an audit opinion, not a warranty, and not proof that any unreviewed system is secure.
Decision · blocked

Findings register decision

Two validated critical findings block retrieval expansion and customer-visible execution until remediation and retest are complete.

Metrics

Findings Register Snapshot

Findings
4
Critical
2
High
2
Validated
4
Release blockers
2
Note

A finding is only useful if it can be fixed and retested

Good red-team findings do not stop at impact. They name the affected boundary, evidence, owner, remediation, and validation standard.

Findings register

Risk register

AI Red-Team Findings Register

The findings register tracks severity, affected boundary, evidence, remediation, validation, and release decision.

Risks
0
Open
0
Critical
0
Decisions
0
Roadmap
0
Controls
0
Data unavailable
content/deliverables/data/ai-red-team-findings-register.json
No risks array found.

Finding summary

Red-team finding summary

IDFindingSeverityBoundaryRelease decision
ART-001Unauthorized retrieval can reach restricted support contextCriticalretrievalblocks retrieval expansion
ART-002Retrieved content can influence model instruction hierarchyHighretrievalrequires remediation before source expansion
ART-003Sensitive customer message action can be queued with thin approval contextCriticalapprovalblocks customer-visible execution
ART-004AI trace schema does not fully reconstruct tool policy decisionsHightracerequires remediation before broad agent expansion

Priority findings

Findings

Priority Findings

Finding · critical

Unauthorized retrieval can reach restricted support context

Evidence: rag-authz-negative-test

Negative tests showed that restricted support context can enter the retrieval candidate set when chunk metadata lacks complete source ACL inheritance.

Finding · critical

Sensitive customer message action can be queued with thin approval context

Evidence: approval-context-test

The system can queue customer-visible response actions without presenting enough evidence for meaningful human approval.

Finding · high

Retrieved content can influence model instruction hierarchy

Evidence: indirect-prompt-injection-test

Synthetic retrieved content containing malicious instructions influenced the model's rationale and answer shape.

Finding · high

AI trace schema does not fully reconstruct tool policy decisions

Evidence: trace-schema-review

Trace records capture tool calls but do not consistently capture policy inputs, denied action class, and approval decision rationale.

Validation criteria

Checklist

Validation criteria before closure

✓Restricted chunks are excluded before retrieval candidates are available.
✓Retrieved instructions are ignored or treated as untrusted context.
✓Approval bundles show target, evidence, rationale, diff, blast radius, rollback, reviewer, and trace reference.
✓Tool traces capture action class, policy inputs, decision, denial reason, target object, and approval state.
✓Release gate blocks expansion until retest passes.
Artifact

Related artifact: AI Red Team Assessment Executive Summary

The executive summary communicates the business decision and release blockers.

/deliverables/ai-red-team-executive-summary
Artifact

Related artifact: AI Red-Team Remediation Roadmap

The remediation roadmap sequences fixes and retest work.

/deliverables/ai-red-team-remediation-roadmap