Skip to content

Security: open-coder-ai/context-report

Security

SECURITY.md

Security Policy

Supported versions

Version Supported
0.1.x (latest) yes
older no

Per in-toto convention, 0.X versions are major: fields may change until 1.0. Only the latest 0.1.x release is supported.

Reporting a vulnerability

context-report defines a report format an author produces in their own CI and a catalog verifies at submission; a flaw here can propagate into every consumer that trusts a statement it produces or verifies.

  • Preferred: open a private security advisory: https://github.com/open-coder-ai/context-report/security/advisories/new
  • Do not open a public issue for exploitable findings.
  • Include: affected schema field, tool path, or spec section; reproduction steps; and impact (e.g. a re-derivable row that is not actually recomputable, a verifier accepting a tampered statement, or a basis a row cannot back up).

Expect an acknowledgment within 7 days and an initial assessment within 14 days. Confirmed vulnerabilities are fixed in a patch release and credited to the reporter unless they prefer otherwise.

Threat model

The spec's central honesty claim is the basis field (re-derivable vs. claimed) and the inputHash that backs a re-derivable row. The two limits below are properties of the design rather than bugs, and are stated here because a report format that leaves them implicit is claiming more than it does.

A statement is only as trustworthy as its verifier

context-report does not itself attest anything; it defines what a producer states and what a verifier is expected to check. A verifier that skips the inputHash recomputation for re-derivable rows, or that treats a claimed row as if it were re-derivable, has silently dropped the one distinction the format exists to make. Nothing in the schema can force a downstream verifier to actually verify.

efficacy rows are claimed, by design

The efficacy row measures paired-ablation lift via a model call (context-report[efficacy], provider: "anthropic" in v0.1). That measurement is stochastic and bound to a model and date — it is a labeled claim, never proof, and is not eligible for basis: re-derivable. A verifier that treats it otherwise is misusing the format, not exploiting a flaw in it.

This repository is itself a chock adopter

Like the organization's other public repositories, this repository governs its own contributions with chock (.agents/policies/). git commit --no-verify skips every git hook, and therefore every compiled gate; the CI-side gate (chock check, chock sync --repo . --check in ci.yml) is the backstop that neither --no-verify nor an unsynced clone can bypass — see chock's own SECURITY.md for the full argument.

There aren't any published security advisories