Skip to content

Open-source agentic AI Search Execution Platform. Track AI visibility, find opportunities, and let agents turn signals into prioritized actions and drafted fixes across ChatGPT, Claude, Gemini, Google AI Overviews, Google AI Mode, Perplexity, Grok, and Microsoft Copilot.

chore(security): flip aggregate/row-fetch RPCs to SECURITY INVOKER (#200)

Gökhan Geyike0b5ab625817
diff --git a/supabase/migrations/00014_security_invoker_rpcs.sql b/supabase/migrations/00014_security_invoker_rpcs.sql
new file mode 100644
index 000000000..728d96e09
--- /dev/null
+++ b/supabase/migrations/00014_security_invoker_rpcs.sql
@@ -0,0 +1,98 @@
+-- 00014_security_invoker_rpcs.sql
+--
+-- Security hardening (defense-in-depth): flip every aggregate / row-fetch RPC
+-- from SECURITY DEFINER to SECURITY INVOKER so the database itself enforces
+-- org isolation via Row Level Security, instead of relying solely on the route
+-- layer to verify brand-belongs-to-org before calling.
+--
+-- Why this is safe (verified against the current schema):
+-- * Every table these RPCs read is already covered by an org-scoped SELECT
+-- policy for the `authenticated` role:
+-- - prompt_results : "Users can read own org prompt results" (00001)
+-- - prompts : "prompts: member select" (00001)
+-- - topics : no RLS, GRANT ALL to authenticated (00001)
+-- * Dashboard callers (web/src/lib/actions/tracking.ts) use the cookie-based
+-- authenticated client, so auth.uid() is always populated and RLS resolves
+-- to the caller's own organization — legitimate numbers are unchanged.
+-- * MCP / worker callers (web/src/lib/mcp/data.ts) use the service_role
+-- client, which bypasses RLS regardless of INVOKER/DEFINER — unaffected.
+--
+-- Effect for a wrong-org call: RLS filters out every row, so aggregates return
+-- zeroed/empty results and row fetches return no rows. No cross-org data leaks.
+--
+-- We use ALTER FUNCTION (not CREATE OR REPLACE) on purpose: it flips only the
+-- security attribute and leaves each function's body, search_path, volatility
+-- and grants byte-for-byte identical — zero risk of body drift.
+
+-- get_latest_prompt_results — two overloads (00001). Currently unused by app
+-- code, but exposed to PostgREST, so harden anyway.
+ALTER FUNCTION public.get_latest_prompt_results(
+ p_brand_id uuid,
+ p_platform text
+) SECURITY INVOKER;
+
+ALTER FUNCTION public.get_latest_prompt_results(
+ p_brand_id uuid,
+ p_platform text,
+ p_model text,
+ p_region text,
+ p_date_from timestamptz,
+ p_date_to timestamptz
+) SECURITY INVOKER;
+
+-- insights_aggregates (current definition: 00012)
+ALTER FUNCTION public.insights_aggregates(
+ p_brand_id uuid,
+ p_platform text,
+ p_models text[],
+ p_region text,
+ p_date_from timestamptz,
+ p_date_to timestamptz,
+ p_prompt_id uuid,
+ p_topic_id uuid
+) SECURITY INVOKER;
+
+-- competitor_aggregates (current definition: 00012)
+ALTER FUNCTION public.competitor_aggregates(
+ p_brand_id uuid,
+ p_platform text,
+ p_models text[],
+ p_region text,
+ p_date_from timestamptz,
+ p_date_to timestamptz,
+ p_prompt_id uuid,
+ p_topic_id uuid
+) SECURITY INVOKER;
+
+-- share_of_voice_aggregates (current definition: 00012)
+ALTER FUNCTION public.share_of_voice_aggregates(
+ p_brand_id uuid,
+ p_platform text,
+ p_models text[],
+ p_region text,
+ p_date_from timestamptz,
+ p_date_to timestamptz,
+ p_prompt_id uuid,
+ p_topic_id uuid
+) SECURITY INVOKER;
+
+-- visibility_trend_aggregates (current definition: 00012)
+ALTER FUNCTION public.visibility_trend_aggregates(
+ p_brand_id uuid,
+ p_models text[],
+ p_region text,
+ p_date_from timestamptz,
+ p_date_to timestamptz,
+ p_topic_id uuid,
+ p_granularity text
+) SECURITY INVOKER;
+
+-- prompt_performance_aggregates (00013)
+ALTER FUNCTION public.prompt_performance_aggregates(
+ p_brand_id uuid,
+ p_models text[],
+ p_region text,
+ p_date_from timestamptz,
+ p_date_to timestamptz,
+ p_topic_id uuid
+) SECURITY INVOKER;