Skip to content

Deployment controls named in reconnaissance can skip the coverage ledger #130

Open

Description

@ornemeolans

What happened

During a full audit of an app backed by a managed Postgres service (Supabase), the reconnaissance agent for deployment visibility correctly reported that the migrations contain no row-level security policies, and that this would matter if the anon key or the auto-generated REST API was exposed.

That note stayed in prose. No coverage unit was seeded for the auto-generated API, and three coverage critics did not flag the gap. Two days later the provider emailed a critical rls_disabled_in_public alert: every public table was readable and writable through the anon role.

A read-only catalog query against pg_tables.rowsecurity and role_table_grants confirmed it in seconds, so the check was cheap and the risk had already been named. It was simply never tracked.

Where the gap is

RECONNAISSANCE.md line 51, Agent 1d, item 4:

Deployed controls and attachments that source cannot establish and therefore require needs_validation if decisive.

This collects the items, but nothing requires each one to enter the coverage ledger, so the parent can drop one without failing any check. Separately, CLOUD-AND-DEPLOYMENT.md has no class for managed backends whose generated API bypasses the application's own authorization layer. The only row-level mentions in the skill are in DATA-ISOLATION-AND-LIFECYCLE.md, and they concern predicates, not default exposure.

Checked against c1c8a8c: the three files cited here match HEAD apart from line endings.

Proposal

  1. Require every Agent 1d item 4 entry to become a ledger unit, either with an owner-observed or read-only check, or as an explicit out_of_scope unit. Never prose only.
  2. Add to the coverage-critic checklist: each recon-listed deployment control has a unit.
  3. Add a class to CLOUD-AND-DEPLOYMENT.md for managed-backend auto-APIs (PostgREST/Supabase, Firebase rules, Hasura), where exposure is the default and application-level authorization isn't in the path.

Related

#73 models how a shared deployment blocker is represented once it exists, and #20 item 4 covers exploitability gating when reachability is a deployment fact. This issue is about the earlier step: a named deployment risk that never reaches the work queue at all.

Activity

  1. ShivamJuyal24 commented on Nov 4, 2026

    @ShivamJuyal24

    Hi, I’d like to take this one.

    Plan, as two small PRs:

    RECONNAISSANCE.md Agent 1d item 4: every entry must become a ledger unit, with a read-only or owner-observed check, or an explicit out_of_scope unit, never prose only. Plus a matching line in the coverage-critic checklist.
    A new CLOUD-AND-DEPLOYMENT.md class for managed-backend auto-APIs (PostgREST/Supabase, Firebase rules, Hasura), where exposure is the default and app-level authorization isn’t in the path.

    I’ve read #20, #73 and PR #74. I see #130 as upstream of both: #74 models how a shared deployment fact is represented in findings.json once it exists, while this issue is about making sure the fact reaches the ledger at all. I’d keep the two independent, with a ledger unit’s check being what a shared_blocker would later reference, and I’d avoid depending on #74 landing first.

  2. ShivamJuyal24 commented on Nov 4, 2026

    @ShivamJuyal24

    Opened #132 for the first part: the recon item 4 requirement and the coverage-critic checklist line. I’ll follow up with the CLOUD-AND-DEPLOYMENT.md managed-backend class in a separate PR once this has feedback.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions