Security at Cogniplane

A governed AI workspace designed for the security review.

Cogniplane is a governed AI workspace for enterprise teams, and the product only works if the security model holds up under real scrutiny. A CISO, an internal auditor, an enterprise procurement team. They all eventually open the architecture and look for the seams. This page is what they find when they do.

Principles

Four things we don't compromise on.

Plenty of AI security stories are written after the fact, once a feature has shipped and someone notices. The four below are constraints we put in front of the product, before it was useful for anything. Each one is enforced in code, in the open-source release you can read yourself, not in a policy document.

01

The model never sees your credentials

Every tool call goes through a per-turn execution context that carries scoped credentials inside the platform. The model gets tool inputs and outputs. It does not get the tokens, headers, or secrets used to authenticate the call.

02

Tenants are isolated by architecture

PostgreSQL row-level security scopes tenant-data queries made through the runtime application role. Sessions are isolated from one another, and shell or file execution runs in a per-session E2B sandbox. Route authorization and database policies both enforce the boundary. Agent checkpoint tables have no RLS, so the runtime reaches them only after a tenant-scoped session check.

03

Approvals are enforced in the runtime

Native runtime approvals gate shell, file, and connector actions, and Policies gate MCP tool calls at the gateway. When either requires approval, execution pauses until the person who started the turn decides. Each Policies approval is bound to the exact request and is used once. The controls live in the platform, not in the prompt, so a user cannot talk the agent past them.

04

The audit trail holds up under review

Sessions, tool calls, approval decisions, and PII findings are all recorded. Tool results are pattern-redacted for secrets before they are written to the audit and tool-event logs, so when a compliance team pulls the trail later, it isn't full of leaked tokens or keys.

Controls

Where each control lives.

If you are running a formal review, this is the map. Each row names a control, what it does, and where in the platform it lives.

01

Authentication

  • WorkOS SSO or local accounts, as enabled for the deployment. Local accounts support passwords, authenticator codes, recovery codes, and passkeys when configured.
  • Local-account MFA can be optional, required for owners and admins, or required for everyone. WorkOS account security is managed by the identity provider.
  • Short-lived JWT access tokens · 15 min
  • httpOnly, Secure refresh cookies that rotate on every use; a replay after the 60-second concurrency grace period revokes the whole token family
  • Redis-backed refresh-token revocation and reuse detection
02

Authorization

  • Role-based access control at the API layer, scoped to owner, admin, member.
  • Agent behavior and Policies enforce built-in tool settings, approval posture, model availability, and tool-level rules inside the runtime execution layer.
  • Projects add their own owner, editor, viewer roles for shared files and conversations.
  • Project chats exclude personal integration accounts, personal-credential MCP connections, and saved memories. The agent cannot read your other chats. Project chats can use organization connections and released project skills. Moving a standalone chat into a project shares its earlier history with project members.
03

Data isolation

  • PostgreSQL row-level security on tenant-data queries made through the runtime application role.
  • Per-session isolation and on-demand E2B code-execution sandboxes prevent cross-session state access.
  • Requests run through a non-superuser role that is itself subject to RLS.
  • Cross-tenant background jobs (scheduler, approval expiry, sign-in lookups) use a separate maintenance role that bypasses RLS but has no superuser, create-database, create-role, or replication rights. Boot checks both roles.
  • The audit log is append-only for the application role: database policies deny updates and deletes.
04

Secret handling

  • Credentials carried in short-lived per-turn execution contexts, never exposed to the model.
  • Tool results pass through automated secret redaction before audit persistence.
  • Tenant API keys are write-only after save.
  • Protected changes to provider keys, data destinations, weaker controls, and shared skill releases require fresh proof from local sign-in sessions. WorkOS sessions do not use this mechanism. Covered changes queue owner notices when email delivery is configured. See confirming sensitive changes.
05

Approval enforcement

  • Admins set native approval posture and Policies rules; both are enforced before execution.
  • Only the user who started the turn can approve it.
  • A Policies approval is bound to the session, user, tool, and a hash of the exact arguments, and the gateway claims it once.
  • Pending approvals expire after 10 minutes by default. The action is rejected, the expiry is written to the audit log, and the user is notified.
  • Scheduled runs cannot wait for approval, so actions that require one are denied.
  • Memory saves and replacements require approval of the exact text. Users can review, edit, and delete saved facts in Settings > Memories.
06

PII protection

  • When enabled, prompts and supported uploads and imports are scanned before use. Supported files include text and PDF text. Optional local OCR reads PDF page images; the verdict records incomplete image coverage.
  • detect reports findings and leaves content unchanged. block rejects detected matches. transform masks chat prompts only; file content stays unchanged.
  • A local detector keeps the detection request local. It does not prevent original content in Detect mode from reaching the agent's model provider. See PII modes and coverage limits.
  • Findings are persisted per message and per artifact for audit. Tenants choose whether matched values are dropped, kept for admins, or kept with per-tenant encryption.
07

Audit trail

  • Every session, tool call, approval decision, and PII finding logged with a replayable history.
  • Per-tenant session review surface in the admin UI.
  • SIEM export — PLANNED
08

Encryption in transit

  • HTTPS everywhere.
  • Cloudflare terminates TLS for the public site and proxies browser traffic to the backend.
  • The public demo reaches its backend through a Cloudflare Tunnel; the host has no public application port.
  • Customer deployments document their internal transport and origin-access controls in the security review.
09

Encryption at rest

  • Storage-level encryption depends on the selected deployment and is documented before production use. Self-hosted operators control their database, object-storage, and volume encryption.
  • Sensitive tenant fields like API keys and integration credentials get a second layer of application-level encryption, using a data-encryption secret.
  • A database read on its own never returns the cleartext value.
10

Retention & backup

  • Backup schedules, off-site copies, restore tests, and recovery targets are deployment-specific production requirements.
  • The public demo is a preview environment. Do not treat it as proof of point-in-time recovery, daily snapshots, or a completed restore drill.
  • Customer retention and offboarding commitments belong in the signed agreement and operating runbook. The tested workflow removes live tenant database rows, unshared stored objects, runtime checkpoints, Redis state, and locally held credentials; third-party revocation and backup expiry follow the deployment-specific agreement.
11

Hardening posture

  • URLs that users supply (skill imports, marketplace manifests) are blocked from private and reserved ranges. Admins can register MCP servers on private networks on purpose; the admin UI warns when they do.
  • Uploads must match a MIME allowlist, and a binary file signature must match the declared type.
  • Download filenames encoded per RFC 5987.
  • Production refuses to start with a default JWT signing secret or a default data-encryption secret.
  • Sandbox outbound access is denied when no hosts are approved. While any host is approved, all traffic to the sandbox's DNS resolver, 8.8.8.8, is also allowed, including DNS over HTTPS. Data can leave through lookup names. The host filter trusts the name stated by the request; it is not a complete data-loss-prevention boundary. Sandbox network settings are separate from MCP Policies.
  • CI fails on any open dependency advisory, and production deploys run their own high-severity audit.
Deployment

How the platform is hosted.

The public demo runs the frontend on Cloudflare Workers. Its backend, PostgreSQL, Redis, and S3-compatible object storage share one self-hosted machine reached through a Cloudflare Tunnel. Shell and file execution runs in isolated per-session sandboxes (E2B), created on demand. The demo proves the application path, not high availability or disaster recovery.

Production topology, data residency, single-tenant isolation, backup, and recovery requirements are agreed before deployment. We will not infer those commitments from the public demo.

Planned · roadmap SIEM export of the audit ledger is on the roadmap, not shipped. Today the audit trail is reviewable per tenant in Admin › Sessions and History, session by session, including every tool call, approval decision, and PII finding. There is no export endpoint yet. If your review needs one, tell us and we will scope it.
FrontendCloudflare
BackendSingle-host demo
IngressCloudflare Tunnel
DatabasePostgreSQL
CacheRedis
ObjectS3-compatible storage
SandboxE2B · code exec · on demand
ReadinessPreview, not HA
ProductionDefined by agreement