pentest gforge loom ai-gateway genai okta aws appsec

GForge — Kickoff Questions

Unknowns to resolve before and during testing. Move items into GForge - Architecture under Confirmed once validated — never silently promote an inference.

Target kickoff: Sep. 9 / 10GForge - Overview > Engagement Timeline


Scope and Targets

To Confirm

  • Exact Loom URL(s)
  • AI Gateway URL(s) / hostnames
  • API base paths
  • Repositories in scope
  • In-scope infrastructure
  • Is the target “Loom + AI Gateway integration”, the gateway service itself, or both plus selected GEL/runtime controls?
  • Which functions are definitely enabled in staging on day one?

Architecture

To Confirm

  • Provide the Confluence architecture diagram referenced in the meeting
  • What exactly is the “Light LLM” component, and where is it deployed?
  • What runtime / orchestration technology is used? (Kubernetes was inferred by one analysis but not confirmed)
  • Where are chat history, projects, user preferences, and any usage/accounting records stored?
  • Do Loom and the AI Gateway have separately reachable endpoints?
  • Is the “Internal Vault” in the secret path the same service as the separately assessed Vault API? → Vault API - Overview

Authentication and Authorization

To Confirm

  • Which Okta token does the gateway validate, and what issuer / audience / scopes / claims are expected?
  • Where is PKCE terminated, and where is token validation enforced?
  • What roles exist (normal user, admin, support, project owner, etc.)?
  • Are Projects and Chat History private, shared, team-based, or organization-wide?
  • Are there multiple approved test identities for isolation testing?

Important

The Projects / Chat History ownership model must be answered before any IDOR / BOLA conclusion.


Gateway and Model Policy

To Confirm

  • Can the client select the model / provider / route directly?
  • Is GLM 5 still the only production-enabled model at test time?
  • Are Kimi / Qwen / self-hosted models enabled in staging?
  • Which Bedrock guardrails or model-governance controls, if any, are currently active?
  • Which budget / rate controls are actually implemented today versus roadmap?

Secrets, Logging, and Data

To Confirm

  • Which secrets are injected into Loom / gateway runtime, and which service identities use them?
  • Where are prompts / responses logged, retained, or observable?
  • Are prompts / responses classified or filtered for sensitive data?
  • What logs are available to testers for correlating security events?

Protocol and Application Behavior

To Confirm

  • Does chat use normal HTTP, streaming SSE, WebSocket, or another transport?
  • Is Swagger / OpenAPI or other gateway API documentation available?
  • Are uploads enabled?
  • Is retrieval / RAG enabled?
  • Are plugins enabled?
  • Is browsing enabled?
  • Are tools / connectors enabled?
  • Is code execution enabled?

Warning

If any of these are not enabled, record them as explicitly excluded rather than leaving scope ambiguous.


Production and Timeline

To Confirm

  • Confirm kickoff target (Sep. 9 / 10)
  • Confirm execution target (week of Sep. 14) and the two-week window
  • Confirm report target (week of Sep. 28)
  • Confirm whether any production parity checks are permitted, and the exact restrictions

Explicit Unknowns — Do Not Assume

Source-listed items that must not be converted into architecture facts.

UnknownNote
Runtime / orchestration platformKubernetes inferred in one analysis, not confirmed by the meeting
“Light LLM” componentExact product/component spelling and implementation unknown
Storage layerDatabase/storage for chat history, projects, preferences, usage/accounting
Authorization / role modelCurrent model in Loom and the gateway
Object ownership semanticsSharing model for Projects and Chat History
Rate limits / quotas / budget controlsExact current values and enforcement
Bedrock guardrailsExact model-governance controls active at test time
Streaming transportExact protocol for streaming responses
Feature enablementWhether uploads, RAG, tools, browsing, code execution, plugins, or connectors are enabled
CLI authentication designFuture design not yet defined