pentest gforge loom ai-gateway genai okta aws bedrock ssm appsec

GForge — Architecture

System understanding for GForge / Loom / AI Gateway. Information is classified by confidence — never treat an inference as architecture fact.

Security controls and boundaries → GForge - Security Model. Scope of what is live → GForge - Current vs Planned Capabilities.


Confidence Legend

LabelMeaning
CONFIRMEDExplicitly stated or demonstrated in the source
PLANNEDDiscussed as future capability — not assumed live
INFERREDReasonable interpretation requiring validation
TO CONFIRMInsufficient evidence

Components

ComponentRoleConfidence
Genesys LoomEmployee-facing chat UI, LibreChat-basedConfirmed
GForge AI GatewayCentral proxy abstracting model accessConfirmed
“Light LLM”Component between the gateway and BedrockConfirmed it exists; exact name/product/deployment To Confirm
AWS BedrockModel hosting / routing backendConfirmed
GLM 5Model stated as the only production-enabled model at meeting timeConfirmed
Okta (PKCE)Browser authentication flowConfirmed
Corporate VPN / VPCInternal-only network placementConfirmed
GELPlatform GForge is described as deployed onConfirmed
Internal Vault → AWS SSM → External Secrets OperatorRuntime secret/config injection pathConfirmed
Runtime / orchestration platformContainer orchestrationTo Confirm — Kubernetes was inferred from “external secrets operator”, not confirmed
Chat history / projects / preferences storageData storesTo Confirm

Current Request Flow

Confirmed

Both source analyses align on this path.

flowchart TD
    U[Internal Employee] --> VPN[Corporate VPN / VPC Boundary]
    VPN --> LOOM["Genesys Loom UI<br/>LibreChat-based"]
    LOOM --> OKTA["Okta Authentication (PKCE)"]
    OKTA --> GW[GForge AI Gateway]
    GW --> LLLM["'Light LLM' component<br/>name TO CONFIRM"]
    LLLM --> BR[AWS Bedrock]
    BR --> M["Approved model<br/>GLM 5 (production-enabled)"]
    M --> LOOM

Step detail:

  1. Internal employee reaches the service from inside the corporate network.
  2. Traffic crosses the corporate VPN / VPC boundary — the service is internal-only.
  3. Genesys Loom UI (LibreChat-based) serves the chat experience.
  4. Okta authentication using PKCE; the user token is presented to the gateway flow.
  5. The GForge AI Gateway receives the request.
  6. The “Light LLM” component sits between the gateway and Bedrock.
  7. AWS Bedrock routes to the approved model.
  8. GLM 5 was the production-enabled model at meeting time.
  9. The model response returns through the same path to Loom.

To Confirm

  • Exactly where PKCE is terminated and where token validation is enforced.
  • Whether Loom and the AI Gateway have separately reachable endpoints.
  • The transport used for chat responses (normal HTTP, streaming SSE, WebSocket, other).

Supporting Secret and Configuration Path

Confirmed

Meeting-established path.

flowchart LR
    V[Internal Vault] --> SSM[AWS SSM Parameter Store]
    SSM --> ESO[External Secrets Operator]
    ESO --> ENV[Application Environment Variables]

This is a supporting dependency, not automatically the primary target. It enters scope only if the approved scope includes the secret-injection boundary.

To Confirm

Whether the “Internal Vault” in this path is the same service as the Vault API assessed separately in Vault API - Overview, or a different internal Vault component. The GForge source does not establish this.

Warning

Avoid turning GForge testing into a standalone Vault pentest. Trace only the approved application-facing portions. Vault has its own engagement — see Vault API - Pentest Game Plan.


Network and Hosting Placement

AspectDetailConfidence
ExposureInternal-only, behind corporate VPC / VPNConfirmed
HostingDeployed on GELConfirmed
OrchestrationContainer platform unspecified — Kubernetes inferred onlyTo Confirm
EnvironmentsStaging environment exists, intended to mirror production as closely as possible for the pentestConfirmed

Important

Internal-only placement is one boundary, not the whole security model. Application controls must still enforce identity and permissions. See GForge - Security Model > Trust Boundaries.


Model Routing

ItemStateConfidence
AWS Bedrock as routing backendCurrentConfirmed
GLM 5 production-enabledCurrentConfirmed (at meeting time)
Kimi 2.5, Qwen 2.5Planned / futureConfirmed as future
Self-hosted / open-source model routesPlanned / futureConfirmed as future
Direct Hugging Face enablementNot plannedConfirmed — the meeting explicitly stated models would be governed/hosted through AWS rather than direct Hugging Face enablement
Bedrock guardrails / model-governance controls activeUnknownTo Confirm

Full capability matrix → GForge - Current vs Planned Capabilities


Source Reconciliation

How the two source analyses compared, per the document.

TopicAssessmentWorking Interpretation
GForge purposeBoth sources alignCentralized internal AI gateway; Loom is a chat product/interface under the initiative
Current request pathBoth sources alignLoom → AI Gateway → Light LLM → AWS Bedrock; Okta/PKCE part of the current auth flow
Network placementBoth sources alignInternal-only, behind corporate VPC/VPN; deployed on GEL
SecretsBoth sources alignInternal Vault → AWS SSM → external secrets operator → application environment
SnykReconciledTranscript says integration is in progress — treat as an active workstream, not necessarily fully deployed
TimelineComplementaryKickoff Sep. 9/10; testing from week of Sep. 14 for ~2 weeks; report week of Sep. 28
KubernetesNot confirmedOne analysis inferred Kubernetes from “external secrets operator”. Keep orchestration technology as To Confirm

Architecture Questions

Open architecture items are tracked in GForge - Kickoff Questions > Architecture. Highest priority:

  • Obtain the Confluence architecture diagram referenced in the meeting
  • Identify what “Light LLM” is and where it is deployed
  • Confirm runtime / orchestration technology
  • Locate storage for chat history, projects, preferences, and usage/accounting records

Sources

  • GForge_AI_Gateway_Loom_Pentest_Gameplan_and_Security_Guidelines.docx (Aug. 31, 2026 meeting analysis and recap).
  • Confluence architecture documentation — to be attached to the ticket, not yet available.