// security

Nothing reaches productionwithout a yes.

How Knotbind handles your Sandbox, credentials and data, and what we watch so a change never breaks in silence. Full data-flow document available to Design Partners.

// how we handle security

Every call goes through the guard.

The assistant never reaches your system directly, and your upstream credential never leaves Knotbind.

Sandbox first

Checks run against your Sandbox only. Every outbound call is routed through the SSRF-guarded client, which allows HTTPS on port 443 to public addresses only — private and link-local ranges (including 169.254.169.254, the cloud metadata service) are refused at the socket. Redirects are followed hop by hop through the same guard, so a hostname cannot pass validation and then re-resolve to a private address (DNS rebinding).

Credentials

Encrypted at rest, write-only in the UI, never logged and never sent to a model. An MCP server's token or API key is sealed with the application secret and exists only for the milliseconds of the real call. The upstream credential is never rendered back to the assistant or to the dashboard, and never written to a Run, Finding or Audit Record.

Scoped keys

AI Assistants hold a scoped Knotbind access key with a monthly budget and a per-key request rate. The upstream credential is never in the conversation there is nothing for a jailbreak to print.

Redaction

Authorization headers, cookies and token, secret and password fields are redacted before traces are stored. When the Content Guardrails module is on, email addresses, US phone numbers, US Social Security numbers and payment card numbers are replaced with [EMAIL], [PHONE], [SSN] and [CARD] in JSON chat message and reasoning text.

Approval gate

No Repair Proposal is applied without explicit Owner Approval. Every Diagnosis is first tested in the Sandbox against read-only fixtures; proposals that fail verification never reach you. Write-capable Checks are never started without that approval.

Audit Record

Every Run, Finding, Diagnosis, proposal and decision is appended to an append-only record that is never edited. A blocked attempt is not silent: an HMAC-signed alert goes to email, Slack or webhook, and the Audit Record keeps the run, the attempt and the decision.

Kill switch

Stop all Runs for an Integration immediately. Revoke credentials at any time. Credentials are operation-bound grants (15–60 minutes) that expire on their own.

EU AI Act compliance & human oversight

The Audit Record and the Owner Approval gate are your enforcers: every assistant action is logged, every fix needs a named person's approval, and no change ships without a yes. Together these satisfy the EU AI Act's human oversight and transparency requirements for high-risk AI systems. Legal compliance remains your decision; Knotbind provides the immutable records and the approval trail.

// what we monitor

Checks that turn failures into findings.

Scheduled watches against your APIs and MCP servers, watching for the changes that break an assistant's answer, and handing you a fix that already passed in the Sandbox.

Contract Checks

Scheduled runs against your OpenAPI description and MCP server. Each tool schema is fingerprinted; a breaking change — a removed tool, a new required parameter, a changed response shape — opens a Finding.

Answer-shape Baselines

The first passing Check records the shape of the answer. A later answer with a different shape fails the Check, so a change nobody announced still surfaces a Finding.

Workflow Checks

Multi-step sequences with output extraction and references between steps. Stop-on-failure by default; each step carries per-step evidence in the Audit Record.

Per-server health

Each MCP server has its own health and uptime, so you can see which connection is down and which is fine, without guessing.

Gateway analytics

When the Request Analytics module is on, every call is logged with its budget and rate usage. Alerts fire on budget overrun, rate-limit breaches, or a model returning an error that the Failover module cannot recover from.

Drift detection

Breaking-drift findings on OpenAPI and MCP schema changes, plus model-output drift when a Baseline no longer matches. Diagnosed and delivered as a tested, Sandbox-verified fix ready for your approval.

// contact

Tell us what your customers should be able to ask.

Send a message. We reply with what to connect, which assistants to target, and a fixed price. No call.