Skip to content

Synthetic Users SDLC AI/GenAI Addendum

Document ID: CRA-13.1.1-AISDLC-001

Version: 1.0

Effective Date: March 25, 2026

Last Updated: March 25, 2026

Owner: CTO — Artur Ventura

Approved By: CTO — Artur Ventura

Classification: Internal – Confidential

CRA Control: CRA 13.1.1

Parent Policy: Secure Software Development Lifecycle (SDLC) Policy v1.6


1. Purpose and Context

This Addendum extends the Synthetic Users Secure Software Development Lifecycle (SDLC) Policy (v1.6, November 2025) to address AI/GenAI-specific development lifecycle requirements mandated by JPMC Control CRA 13.1.1. It supplements — but does not replace — the parent SDLC policy, adding binding requirements wherever AI or Generative AI systems are involved.

JPMC CRA 13.1.1 requires confirmation that a documented SDLC process covers AI/GenAI models, encompassing model validation, risk assessments, policies for retraining models, and testing strategies. This Addendum also formally documents the company's alignment with the CWE/SANS Top 25 vulnerability standard across all developed applications.

1.1 Why This Addendum Is Needed

Synthetic Users' platform incorporates AI/GenAI capabilities including a multi-provider LLM orchestration layer (LLM Shuffle), a Persona Engine based on the OCEAN personality model, and a Retrieval-Augmented Generation (RAG) pipeline. These capabilities introduce lifecycle considerations not fully addressed by a traditional Secure SDLC:

  • Non-deterministic model outputs requiring specialised testing approaches
  • Third-party model vendor risks (prompt injection, data exfiltration via outputs)
  • Data flow concerns unique to prompts, context windows, and inference logs
  • Model selection, validation, and deprecation cycles distinct from traditional software releases
  • Regulatory and contractual AI disclosure obligations, including JPMC CRA 13.1.1

1.2 Scope

This Addendum applies to all AI/GenAI features in the Synthetic Users production platform, all LLM providers accessed via API, all engineers and contractors involved in AI/GenAI development, and all RAG components, prompt templates, and agent workflows.

1.3 Exclusions

This Addendum does not apply to internal productivity AI tools (e.g., GitHub Copilot) unless those tools process customer data or JPMC data.


2. CWE/SANS Top 25 Alignment

Synthetic Users' SDLC explicitly incorporates the CWE/SANS Top 25 Most Dangerous Software Weaknesses as a vulnerability taxonomy for all developed applications, in addition to the OWASP Top 10. This section formally documents that alignment.

2.1 Integration in SAST and DAST Workflows

  • SAST tooling (GitHub Advanced Security / CodeQL) is configured to report CWE IDs alongside findings
  • DAST results include CWE references in remediation notes provided to engineering
  • Penetration test reports are required to map findings to CWE identifiers where applicable
  • Vulnerability management SLAs apply to all CWE/SANS Top 25-mapped findings

2.2 Key CWE/SANS Top 25 Weaknesses — Platform Mapping

CWE IDWeaknessRelevance to Synthetic UsersMitigation Control
CWE-79Cross-site Scripting (XSS)Web app renders dynamic AI output contentOutput encoding, CSP headers, LLM response sanitisation
CWE-89SQL InjectionDatabase queries for research sessionsParameterised queries (Prisma ORM); SAST injection checks
CWE-20Improper Input ValidationPrompt inputs; JPMC API inputsInput length limits, schema validation, prompt boundary controls
CWE-352CSRFState-changing API callsCSRF tokens; SameSite cookies; Google Firebase session controls
CWE-22Path TraversalRAG document ingestion pipelineAllowlist-based path validation; sandboxed ingestion workers
CWE-732Incorrect Permission AssignmentMulti-tenant data isolationRBAC; row-level security; tenant ID propagation checks
CWE-502Deserialization of Untrusted DataLLM output parsing; API response handlingType-safe parsers; schema validation
CWE-862Missing AuthorisationAPI endpoints for session managementMiddleware auth checks; automated endpoint coverage testing
CWE-798Hard-coded CredentialsLLM provider API keysRender environment variables; 1Password; SAST secret detection
CWE-287Improper AuthenticationUser login; admin access; JPMC API authGoogle Firebase (application auth); Google Workspace (employee auth); MFA enforcement; API key rotation policy

2.3 Remediation SLAs

SeverityDefinitionSLAEscalation
CriticalCWE/SANS Top 25 + active exploitation or CVSS ≥ 9.024 hoursCTO + CEO immediate notification
HighCWE/SANS Top 25 + CVSS 7.0–8.97 calendar daysCTO notified; tracked in sprint
MediumCWE/SANS Top 25 + CVSS 4.0–6.930 calendar daysEngineering lead tracks
LowCVSS < 4.0 or informational90 calendar daysTracked in backlog

3. AI/GenAI Model Lifecycle

AI/GenAI systems at Synthetic Users follow a structured lifecycle that integrates with the parent SDLC. Unlike traditional software, AI/GenAI features involve the selection, validation, integration, monitoring, and retirement of third-party language models — not the training of models from scratch.

PhaseActivitiesResponsibleGate Criteria
1. Model Selection & FeasibilityIdentify candidate LLM providers via LLM Shuffle evaluation; assess capabilities, data handling terms, security postureCTO / Engineering LeadProvider must pass DPA review; no customer PII sent without DPA
2. Security & Privacy ReviewReview provider's data retention, logging, fine-tuning opt-outs, incident response SLAs; prompt injection risk assessment; DFD updateCTO + LegalSigned DPA; DFD updated; no customer data logged without opt-in
3. Integration & DevelopmentImplement API via LLM Shuffle; develop prompt templates; build RAG connectors; apply output sanitisationEngineeringCode review complete; SAST passing; no hardcoded keys
4. AI-Specific TestingExecute AI/GenAI test suite; validate output quality, safety, security; conduct adversarial prompt testingEngineering + QATest suite passing; prompt injection scenarios tested and mitigated
5. Staged DeploymentDeploy to staging; canary release to production subset; monitor output qualityEngineeringCanary metrics within bounds; rollback procedure tested
6. Production MonitoringOutput quality monitoring; anomaly detection; latency tracking; periodic adversarial re-testingEngineering + CTOAutomated alerts active; AI/GenAI incident playbook in place
7. Model RetirementPlan migration; update prompt templates; re-run test suite; confirm data deletion from retired providerEngineering + LegalMigration tests passing; provider data deletion confirmed in writing

4. Model Validation Framework

Before any AI/GenAI model or provider is enabled in production, it must pass a structured validation process.

4.1 Validation Checklist — New Model or Provider

CategoryRequirementEvidence
Data HandlingProvider does not train on or retain API request data by defaultWritten confirmation via provider DPA
Data HandlingNo customer PII or JPMC data sent to provider without DPA coverageDFD review; legal sign-off
SecurityProvider enforces TLS 1.2+ in transit; at-rest encryption confirmedProvider security docs or SOC 2 report
SecurityPrompt injection risk assessment completedInternal assessment document; adversarial test cases
SecurityAPI key lifecycle defined: rotation frequency, storage in Render env varsSAST scan; 1Password entry confirmed
QualityModel output evaluated against representative test setEvaluation report; pass/fail criteria documented
QualityOutput hallucination and safety behaviour assessedTest report; content policy acceptance recorded
ContractualProvider terms permit B2B commercial useLegal review sign-off
Incident ResponseProvider's incident response SLA and notification obligations confirmedDPA or security exhibit reference
ResilienceFallback provider identified in LLM Shuffle configurationLLM Shuffle fallback config documented

4.2 Validation for Model Capability Changes

When a provider releases a significant model update, a lightweight re-validation is conducted:

  • Review of provider change notes for data handling or safety behaviour changes
  • Re-run of core AI test suite against the updated model endpoint
  • Comparison of output quality metrics against baseline
  • CTO sign-off before switching default model in LLM Shuffle

5. AI/GenAI Risk Assessment

AI/GenAI features are subject to a dedicated risk assessment in addition to standard threat modelling.

5.1 When Required

  • Introduction of a new LLM provider or model to production
  • New AI/GenAI feature processing user-provided data or JPMC data
  • Changes to prompt templates affecting data flow
  • Changes to the RAG pipeline (ingestion sources, retrieval logic)
  • Changes to output post-processing or content filtering

5.2 Required Topics

Risk AreaKey QuestionsTypical Controls
Prompt InjectionCan a malicious user override system instructions or exfiltrate context window data?Input validation; prompt boundary enforcement; adversarial test cases
Data Leakage via OutputCan the model output other tenants' data or RAG index content it shouldn't expose?Tenant ID scoping in RAG; output PII scanning; session isolation
Hallucination RiskCould incorrect outputs cause harm in synthetic user research (false attribution, fabricated personas)?AI-generated labelling; downstream validation; user disclaimers
Third-Party Model RiskWhat are the provider's retention, logging, and training data practices?DPA review; training opt-out; API log review
Availability & ResilienceWhat is the impact of LLM provider downtime? Is there a fallback?LLM Shuffle fallback; circuit breaker patterns; degraded-mode UX
Regulatory ComplianceDoes the AI feature trigger GDPR, CCPA, or JPMC contractual obligations?DPIA if required; data minimisation; consent; JPMC notification
Model MisuseCould the AI feature be used to generate harmful or deceptive content?Content filtering; system prompt scoping; rate limiting; abuse detection

6. Retraining and Model Update Policy

6.1 Current Architecture — No Fine-Tuning

Synthetic Users does not fine-tune or retrain base language models. All AI/GenAI capabilities are delivered through third-party LLMs via API, with context provided at inference time via RAG. This architecture materially reduces model retraining risk.

Because no proprietary model weights are maintained, "retraining" in the traditional sense does not apply. Instead, the company manages prompt template versioning, RAG index updates, and LLM provider version management as the functional equivalents.

6.2 Prompt Template Versioning

  • All prompt templates are stored in the version-controlled repository under /prompts/ with semantic versioning
  • Changes to prompt templates affecting data flow or output behaviour require engineering review and testing
  • Prompt template releases follow the same CI/CD pipeline as code changes
  • Breaking changes to system prompts require CTO approval

6.3 RAG Index Update Policy

  • New document sources added to the RAG index require security review
  • RAG index contents are scoped by tenant — no cross-tenant retrieval is permitted
  • JPMC data is not ingested into the shared RAG index; JPMC context is provided in the prompt context window only
  • Index integrity checks are run after bulk updates

6.4 Model Provider Version Management

TriggerActionApproval
Provider deprecates model version in productionMigration plan created ≥30 days before deprecation; test suite re-runEngineering Lead + CTO
Provider releases major new versionLightweight validation; AI test suite re-run; output comparisonEngineering Lead; CTO for default switch
Provider changes data handling policyLegal and CTO review; re-validate DPA; update DFDCTO + Legal
Security vulnerability in model or provider APIImmediate risk assessment; consider fallback via LLM ShuffleCTO — immediate escalation

7. AI/GenAI Testing Strategies

7.1 Required Test Types

Test TypeWhat It ValidatesFrequencyTooling
Prompt Regression TestingPrompt templates produce expected output structure and content for fixed representative inputsEvery deployment modifying a prompt templateCustom test harness; CI/CD assertions
Adversarial Prompt TestingSystem resists prompt injection, jailbreak, and instruction overrides; no context window leakageBefore each new model integration; quarterlyManual red-team + automated injection pattern library
Output Safety TestingOutputs are free of harmful content; AI-generated content is correctly labelledBefore each model or system-prompt changeContent policy checklist; manual output review
Data Isolation TestingTenant A's RAG context is not retrievable by Tenant B; no cross-tenant leakageBefore each RAG pipeline change; quarterlyMulti-tenant test accounts; retrieval audits
Hallucination & AccuracySynthetic persona outputs accurately reflect input parametersPer release cycle; sampled in productionGolden-set comparison; human review
Performance & LatencyLLM response times meet SLA; fallback provider activation testedLoad tested quarterly; canary phasek6; LLM Shuffle failover drill
Integration SecurityLLM API auth, key rotation, error handling correct; no keys in logsEvery deployment; annual pentest scopeSAST (CodeQL); code review; pentest

7.2 OWASP Top 10 for LLM Applications — Alignment

OWASP LLM RiskSynthetic Users Control
LLM01 — Prompt InjectionInput validation; system prompt boundary enforcement; adversarial test cases in CI/CD
LLM02 — Insecure Output HandlingOutput sanitisation before rendering; HTML encoding; CSP headers; no raw LLM output to DOM
LLM03 — Training Data PoisoningNot applicable (no fine-tuning); RAG index integrity checks and source validation
LLM04 — Model Denial of ServiceRate limiting on AI endpoints; timeout and retry logic; LLM Shuffle fallback
LLM05 — Supply Chain VulnerabilitiesLLM provider security reviews; DPA validation; Dependabot on AI SDK dependencies
LLM06 — Sensitive Information DisclosureRAG tenant scoping; no PII in shared prompts; output review for data leakage patterns
LLM07 — Insecure Plugin DesignAll LLM tool/function calls reviewed for injection risk; allowlist of permitted actions
LLM08 — Excessive AgencyAgent permissions constrained; human-in-the-loop for high-risk actions
LLM09 — OverrelianceOutputs labelled as AI-generated; downstream validation steps documented
LLM10 — Model TheftAPI keys rotated regularly; access logs monitored; no model weights stored internally

8. SDLC Toolchain Security

Access to the AI/GenAI SDLC toolchain is controlled with the same rigour as general software development:

Toolchain AssetAccess ControlPrivileged Access
GitHub repository (prompts/, AI code)RBAC; PR review required for all mergesAdmin rights limited to CTO; engineers via SSO + MFA
LLM provider API keysStored in Render env vars and 1Password; never in codeRotation requires CTO approval; CI/CD uses scoped service keys
RAG index / vector storeTenant-scoped; no direct production access outside break-glassBreak-glass requires CTO authorisation; fully logged
AI monitoring dashboardsRead-only for engineers; alert config limited to Engineering LeadSSO enforced; no shared credentials
LLM provider admin portalsCTO-only access to billing, usage analytics, and model managementMFA enforced; access reviewed quarterly

9. Roles and Responsibilities

RoleResponsibilities
CTO — Artur VenturaPolicy owner; approves new model integrations; approves breaking changes to system prompts; leads AI/GenAI risk assessments; escalation for AI security incidents
Engineering LeadImplements and enforces this Addendum; owns AI test suite; coordinates adversarial testing; reviews model validation checklists
EngineersFollow AI/GenAI coding standards; include adversarial test cases in PRs; flag anomalous model behaviour; never hardcode API keys
Legal / ComplianceReviews and approves DPAs for new LLM providers; advises on JPMC data handling; reviews model retirement data deletion confirmations
CEO — Kwame FerreiraReceives immediate notification of Critical AI/GenAI security incidents; approves strategic AI provider decisions

10. Evidence and Audit Requirements

Evidence ArtefactCreated WhenRetentionStorage
Model Validation ChecklistEach new model or provider integration3 yearsInternal compliance folder
AI/GenAI Risk AssessmentEach material AI feature change3 yearsInternal compliance folder
Prompt Template Change LogEach prompt template release3 yearsGitHub repository history
AI Test Suite ResultsEach relevant deployment1 year rollingCI/CD pipeline artefacts
Adversarial Test ReportsQuarterly + per new model integration3 yearsInternal compliance folder
RAG Index Audit LogsContinuous1 year rollingAWS CloudWatch / logging service
Provider DPA RecordsEach provider engagementContract + 3 yearsLegal folder
Model Retirement RecordsEach model deprecation or provider exit5 yearsLegal + compliance folder

11. Change Management

VersionDateAuthorSummary
1.0March 25, 2026Artur Ventura, CTOInitial release. Establishes AI/GenAI lifecycle, model validation, risk assessment, retraining policy, CWE/SANS Top 25 alignment, and AI-specific testing strategies as mandated by JPMC CRA 13.1.1.

This Addendum is reviewed annually and whenever a material change occurs to the AI/GenAI platform, LLM provider roster, or applicable regulatory requirements.



Synthetic Users, Inc. — 3201 Coolidge Ave, Los Angeles, CA 90066, USA

Released under the MIT License.