insecur is experimental and prelaunch. There is no public sign-up yet, and the hosted service is not approved for valuable production secrets. Local Mode runs the development loop with no account.

Security model

insecur is experimental and prelaunch. This page separates the development workflow available today from the protected-environment design that still needs launch proof. Do not use the hosted service for valuable production secrets yet.

The guarantee differs by tier

Development environments Protected-environment design
Plaintext at rest Never Never
Value reaches your machine Yes, injected into one child process per grant No
Read-back through the product No response body contains a value No read path
Delivery requires A logged-in session and a one-use grant An environment-bound machine credential
Change control Writes go live immediately Draft versions, promotion, human approval
Readiness Experimental development workflow Prelaunch and not launch-proven

For development, an agent-launched process holding the injected value is a process the agent controls, so the agent can read or print that value. The development benefit is no plaintext project file to scrape, no standing credential in a shell profile, and one audited grant per use. For the protected-environment design, the readable value does not reach an agent-reachable machine, and no human session, CLI call, or agent channel can obtain a protected injection grant. That stronger path is not launch-proven.

Structural isolation, not conditionals

The system runs as capability-isolated deploys. The public API edge holds no decryption capability and no database binding. The private runtime service is the sole holder of the root key binding, the only place decryption happens, and serves zero public routes; it is reachable only over a private service binding. The web console and the public site hold neither. CI conformance gates fail the build if any deploy ever combines a public route with the root key, so the boundary cannot erode quietly in a refactor.

Cryptography

Secret values are protected with AES-256-GCM envelope encryption under a tenant-bound key hierarchy: an instance root key wraps organization and project data keys, which wrap per-record data keys. Ciphertext is bound to its identity, so moving ciphertext between records fails closed. Key custody and rotation procedures harden further before general availability; current state is always reflected on the security page.

Input hygiene the CLI enforces

  • Secret values are never accepted as command-line arguments. Inputs are stdin (--value-stdin), masked prompts, service-side generation (--generate), or provider authorization. --token style flags are explicitly rejected.
  • All CLI and API output is metadata-only by contract, enforced by shape checks and a no-plaintext canary test suite in CI.
  • Crash reports and telemetry are sanitized metadata: never argv, environment, values, or transcripts.

Accountability

Every meaningful action writes a tenant-scoped audit event carrying the principal chain: the human, the agent session acting under them, or the machine identity. Exports are signed (Ed25519) and independently verifiable against published keys. The claim ceiling is tamper-evident and independently verifiable, nothing stronger.

What we do not claim

  • Not zero-knowledge. We do not claim the operator is technically incapable of accessing data. Operator access is constrained by controls and audit, and the trust model is published rather than hand-waved.
  • No secrecy from your own runtime. A workload that receives a secret can do anything a process can do. insecur controls delivery and evidence, not what your code does afterward.
  • Local Mode is encrypted local custody, not no-reveal custody. The machine key lives on the same machine.
  • No secure-erasure claims. insecur local-files rm is an ordinary filesystem delete.
  • No automatic provider rotation or one-button exposure recovery today. Writing a replacement value to insecur does not revoke the old credential at its provider.

Reporting

Found a vulnerability? See the security page for the current disclosure contact. Please do not test against other tenants' data.