Exposed Model API Keys
Putting OpenAI or Bedrock credentials in agent pods risks key theft during a prompt injection. How Zelkor stops it: Provider keys stay in gateway Kubernetes Secrets; agent pods hold only a local gateway token.
Deploy LangGraph, LangChain, and Deep Agents workloads on Kubernetes with zero model keys in agent pods, gVisor isolation for quantitative Python scripts, read-only tenant-scoped Postgres queries, and self-hosted Langfuse traces.
Putting OpenAI or Bedrock credentials in agent pods risks key theft during a prompt injection. How Zelkor stops it: Provider keys stay in gateway Kubernetes Secrets; agent pods hold only a local gateway token.
Running generated portfolio math inside the agent container exposes the pod’s filesystem and service account. How Zelkor stops it: Untrusted Python runs over MCP in a warm pool of gVisor worker pods.
Shared research graphs queried by multiple desks risk returning another client’s holdings. How Zelkor stops it: Verified JWT identity sets the tenant session for Postgres Row-Level Security and forces tenant filters on Qdrant.
Hallucinated SQL statements attempting to modify or delete core financial records. How Zelkor stops it: The Postgres MCP tool enforces read-only queries.
Recursive agent loops firing hundreds of model calls per minute. How Zelkor stops it: The AI gateway enforces a per-minute request cap and rejects extra calls before reaching the provider.
Standard container logs capturing sensitive account numbers and prompts. How Zelkor stops it: Default log settings omit prompt bodies, bearer tokens, and API keys.
| Requirement | What Zelkor Does | How You Check It |
|---|---|---|
| Network Segmentation (DORA / PCI-DSS) | Kubernetes NetworkPolicy blocks agent pods from calling the internet, sandbox workers, or databases directly. | Run The Vanishing Key and The Shell Moved network proofs. |
| Credential Custody & Access (SOC 2) | Keeps upstream LLM keys on the gateway and rejects unsigned or invalid JWTs at the front door. | Inspect pod environment variables and run negative JWT tests. |
| Multi-Desk Information Barriers | Propagates verified user identity to MCP tools, sets Postgres tenant sessions for RLS, and filters Qdrant by tenant. | Run Two Desks, One Graph cross-tenant proof. |
| Model Risk Audit Trail (SR 11-7) | Records full run waterfalls and automated run checks in self-hosted Langfuse; inspects prompts with NeMo Guardrails. | Review run traces in Langfuse via The Flight Recorder proof. |
Serve equity and credit research across separate desks from one LangGraph deployment while enforcing desk-level isolation in Postgres and Qdrant.
Execute agent-written Python valuation models inside the warm gVisor worker pool with no access to the agent pod or external networks.
Route every chat completion through fail-closed NVIDIA NeMo Guardrails for topic safety and cap request rates at the cluster gateway.