RESEARCH / SPECIFICATION

The CDK Threat Model

Version 1.0 Published 2026-07-29 Authors CDK Core Team DRAFT

Abstract

This document enumerates the threats CDK's evidence model is designed to resist, the assumptions it depends on, and what remains explicitly out of scope. It complements the Security Architecture, which describes the mechanisms; this document describes what those mechanisms are for — and where they stop.

1. Scope

This document covers threats to the integrity of a single attestation and the Trust Graph built from it. It does not cover the security of the systems that produce the underlying events — CDK observes and attests; it does not harden the CI/CD platforms, cloud infrastructure, or identity providers it sits alongside.

2. Threats & Mitigations

T1

Attestation Forgery

An attacker creates a fake attestation claiming an event occurred that didn't.

MITIGATION

Every attestation requires a valid Sigstore signature tied to a verified OIDC identity. Forged attestations fail verification immediately.

T2

Evidence Tampering

An attacker modifies a stored attestation after the fact.

MITIGATION

Hash-chained storage breaks verifiably if any record changes. Rekor's independent log provides a reference point outside CDK's own database.

T3

Replay

A valid, old attestation is resubmitted to claim a new event occurred.

MITIGATION

Every attestation is bound to a specific artifact digest and timestamp. Replayed attestations reference stale digests and fail policy checks expecting current ones.

T4

Identity Spoofing

An attacker impersonates a legitimate approver or signer.

MITIGATION

Signing requires a live OIDC token from the identity provider itself. CDK never accepts a bare claim of identity without cryptographic proof behind it.

T5

Compromised CI Runner

A build system is compromised and produces a legitimate-looking but malicious artifact.

LIMITATION, NOT MITIGATION

This is explicitly out of scope for the Evidence Model alone — CDK attests to what a system reported, not whether that system was trustworthy. The digest binding still matters here: a compromised build produces a verifiable record of exactly what shipped, enabling detection after the fact even though it can't be prevented beforehand.

3. Assumptions

OIDC Provider
Assumed trustworthy — its tokens are not compromised or issued for spoofed identities.
Sigstore Infrastructure
Fulcio and Rekor are assumed available and uncompromised at time of signing.
Source Platform Access
The CI/CD platform enforces its own basic access controls; CDK does not replace them.

4. Out of Scope

  • Securing the CI/CD platform itself
  • Vetting the identity provider's own security practices
  • Preventing a compromised build from producing a malicious artifact — only ensuring its provenance stays traceable afterward

References

The CDK Security Architecturev1.0
The CDK Evidence Model Specificationv1.0
Sigstore — Keyless Signingsigstore.dev