pentest gforge loom ai-gateway genai okta aws bedrock appsec api-security

GForge — Overview

Entry point for the GForge / AI Gateway / Loom engagement. This note is the dashboard — detail lives in the linked notes.


What These Systems Are

TermWhat It Is
GForgeInternal initiative for centralized access to LLMs — the broader programme
LoomCurrent user-facing product under GForge; a ChatGPT-like internal chat interface based on LibreChat
AI GatewayCentral proxy layer that abstracts model access and is intended over time to centralize governance (model selection, budgets, rate limits)

Note

GForge is the initiative. Loom is one current product under it. The AI Gateway is the service that actually proxies model requests. Keep the three distinct when scoping.

GForge in one sentence (from source): an internal AI gateway ecosystem where Loom currently provides the employee chat experience and the gateway proxies authenticated requests toward approved AWS-hosted model backends.


Current Request Flow

Confirmed

Both source analyses align on this path.

flowchart TD
    A[Internal Employee] --> B[Corporate VPN / VPC Boundary]
    B --> C["Genesys Loom UI<br/>(LibreChat-based)"]
    C --> D["Okta Authentication (PKCE)<br/>user token presented to gateway flow"]
    D --> E[GForge AI Gateway]
    E --> F["'Light LLM' component<br/>(exact name TO CONFIRM)"]
    F --> G[AWS Bedrock]
    G --> H["Approved model<br/>GLM 5 enabled in production"]
    H --> C

To Confirm

The exact product name, spelling, and deployment location of the “Light LLM” component must be confirmed from the repository or architecture documentation.


Supporting Secret / Configuration Path

Meeting-established path:

Internal Vault → AWS SSM Parameter Store → External Secrets Operator → application environment variables

This is a supporting dependency of GForge/Loom. It is not automatically the primary target unless the approved scope includes the secret-injection boundary.

Detail → GForge - Architecture > Supporting Secret and Configuration Path · Related system → Vault API - Overview


Core Security Question

Important

Can an authenticated internal user make Loom or the AI Gateway perform actions, access data, invoke models/routes, or expose information beyond what that user and the approved application workflow are intended to permit?


Major Trust Boundaries (Summary)

BoundaryWhy It Matters
Corporate VPN / VPCNetwork placement — internal-only, but not a substitute for app-layer controls
Okta / PKCE authenticationIdentity establishment in the browser flow
Token validation at the gatewayWhere identity must actually be enforced
Loom application authorizationOwnership of chats, projects, history
AI Gateway policy enforcementApproved models, providers, routes
Model context boundarySecrets and system context must not leak into prompts/responses
Secret injection pathRuntime secrets must not reach the browser, logs, or model context
Environment separationStaging vs production parity and data isolation

Full detail → GForge - Security Model


Current vs Planned (Summary)

CurrentPlanned / Future
Loom / LibreChat web UIBudget enforcement
AI Gateway model proxyRate-limit governance
AWS Bedrock routing (GLM 5)CLI access + separate auth flow
Okta + PKCE authenticationAgentic email / Teams / GitHub workflows
Internal-only VPN/VPC placementAdditional / self-hosted models
Vault → SSM → secret injectionBroader model governance protocols
Snyk integration (in progress)

Warning

Do not build active test cases for planned capabilities unless they are actually enabled and approved in the test environment.

Full matrix and scope guardrails → GForge - Current vs Planned Capabilities


Engagement Timeline

Confirmed

Targets stated in the source. Confirm at kickoff.

MilestoneTarget
KickoffSep. 9 / 10
Testing executionWeek of Sep. 14, approximately two weeks
ReportWeek of Sep. 28

Note

Vault testing was described as taking priority over GForge. See Vault API - Pentest Game Plan.


Important Unknowns

Highest-impact open items — full list in GForge - Kickoff Questions.

  • Exact Loom and AI Gateway URLs, hostnames, and API base paths
  • What the “Light LLM” component actually is
  • Runtime / orchestration platform (Kubernetes inferred, not confirmed)
  • Authorization and role model in Loom and the gateway
  • Ownership / sharing semantics for Projects and Chat History
  • Whether the client can select model / provider / route directly
  • Current rate limits, quotas, and budget controls
  • Storage location for chat history, projects, preferences, usage records
  • Transport for streaming responses (HTTP, SSE, WebSocket, other)
  • Whether uploads, RAG, tools, browsing, plugins, or code execution are enabled

Relationship to GEL and Vault

Keep the systems separate during scoping.

SystemRoleScope Note
GForge / LoomAI service / application under assessmentPrimary target
GELHosting / platform context — GForge is described as deployed on GELGEL itself is not automatically in scope
VaultSeparate internal secrets service feeding AWS SSMAlready a distinct pentest target
AWS SSMSecret / config backend consumed by the external secrets operatorSupporting path
SnykAutomated security input; repository integration in progressTriage input

GEL - Overview · Vault API - Overview · Genesys Security Architecture - Overview


Note Map

NoteUse During Engagement
GForge - ArchitectureComponents, request flow, trust-boundary detail, confidence labels
GForge - Security ModelAuthentication, authorization, model routing, secrets, AI security principles
GForge - Current vs Planned CapabilitiesScope guard — what is live vs roadmap
GForge - Pentest Game PlanObjectives, frameworks, phases, priorities, guardrails
GForge - Endpoint InventoryWorking table of Loom / gateway endpoints
GForge - Test CasesTest themes and checklists by category
GForge - Evidence LogChronological observations with model/build context
GForge - FindingsCandidate and confirmed findings only
GForge - Kickoff QuestionsUnknowns to resolve before and during testing

Sources

  • GForge_AI_Gateway_Loom_Pentest_Gameplan_and_Security_Guidelines.docx — working security-testing reference based on the Aug. 31, 2026 meeting analysis and recap.
  • Architecture documentation was to be attached to the ticket from Confluence (not yet received).