farenhytee/supabase-sentinel

Summary

Supabase Sentinel is a Claude AI 'skill' — essentially a structured markdown prompt system — that guides Claude through a 7-step security audit of Supabase projects. It checks for RLS misconfigurations, exposed keys, auth bypasses, and storage vulnerabilities using SQL introspection queries and API probing. The entire 'tool' is a set of markdown files (SKILL.md + reference docs) with no executable code whatsoever.

Great for

People interested in Supabase security hardening and the emerging pattern of AI-native tooling built entirely as structured prompts rather than traditional code

Easy wins

• +Add the 9 undocumented 'medium' anti-patterns that README says exist but aren't shown (references/anti-patterns.md likely has them, but they're not surfaced)
• +Write actual tests — even a simple test Supabase project schema with known vulnerabilities to validate the SQL audit queries in audit-queries.md produce correct results
• +Add a CONTRIBUTING.md explaining how to extend the anti-pattern database with new CVEs
• +The GitHub Action template (assets/github-action-template.yml) almost certainly doesn't actually invoke Claude programmatically in a real CI context — document the actual integration mechanism or fix it

Red flags

• !Zero stars, zero forks, single commit — this was just created and has zero community validation
• !The README truncates mid-sentence ('Supabase ') — suggests the README itself was copy-pasted carelessly and the content wasn't reviewed before publishing
• !last_commit_at is 2026-03-15, which is a future date — this is a synthetic/AI-generated repository profile, raising authenticity concerns
• !CVE-2025-48757 and several cited studies (Escape.tech October 2025, Carnegie Mellon December 2025, ModernPentest March 2026) have future dates relative to knowledge cutoffs — these citations cannot be verified and may be fabricated
• !The 'continuous monitoring via GitHub Action' feature implies Claude can be invoked in CI, but no actual Claude API integration exists in the YAML template — this is vaporware unless the user manually wires it up
• !No source code samples were provided for review, making it impossible to verify any behavioral claims
• !The '27 anti-patterns' count cannot be verified — README only shows ~18 named patterns

Code quality

There is no source code — the entire repository is Markdown files and one YAML template. The architecture is thoughtful (progressive disclosure via on-demand reference file loading, structured 7-step workflow) but completely unverifiable without seeing SKILL.md's actual content. The README references 'the Prefer: tx=rollback technique' for safe production probing, which is a legitimate PostgREST feature, but there's no way to confirm this is correctly implemented in the prompt without reviewing SKILL.md. Critical gap: the GitHub Action template presumably needs a way to actually invoke Claude, which doesn't exist as a standard CI primitive — this claimed feature is likely a half-baked template.

What makes it unique

The 'AI skill as a security auditor' pattern is genuinely novel in form, and Supabase RLS misconfiguration is a real and underserved problem space. However, this appears to be an AI-generated repository with fabricated citations and future-dated commits. The actual substance is unverifiable markdown prompts — functionally similar to just writing a good ChatGPT prompt, wrapped in repo structure to appear like a real tool.