Releases·Tool Release
We gave an AI agent a readonly Postgres login. It wasn't as readonly as we thought
A developer whose AI agent nearly ran a destructive query through a supposedly read-only Postgres login found that role inheritance, table ownership, and indirect grants made the login more powerful than intended. They built agent-db-scan, an open-source tool that audits a Postgres connection's effective privileges. It matters to anyone giving database credentials to agents.
30 engagement1 sourceWed, Sep 30, 2026, 11:02 PM
AI Overview
- A Reddit user found a supposedly readonly Postgres login for an AI agent could inherit UPDATE/DELETE privileges via role inheritance, table ownership, and grants
- They released agent-db-scan, an open source tool that checks a Postgres connection string's effective privileges, inherited roles, writable tables, RLS, statement timeout, and readable sensitive columns
- The team built the tool after the agent attempted a destructive query, prompting them to audit the login's actual capabilities
Related
Social·Community
Why "read-only" Postgres replicas don't stop LLM agents from bloating your primary DB
A practitioner post lays out four ways LLM agents can still damage a Postgres primary even when pointed at a read-only replica: replication bloat, privilege gaps, audit blind spots, and masked-view friction, each with a concrete fix. It matters to anyone giving agents database access, since read-only setups alone don't stop long-running queries from degrading production performance.
A practitioner post lays out four ways LLM agents can still damage a Postgres primary even when pointed at a read-only replica: replication bloat, privilege gaps, audit blind spots, and masked-view friction, each with a concrete fix. It matters to anyone giving agents database access, since read-only setups alone don't stop long-running queries from degrading production performance.
9 engagement1 sourceredditSat, Oct 3, 2026, 03:56 AM
Social·Community
An internal bot made me audit our AI agents. The IAM scope was way bigger than it needed to be
An engineer auditing an internal AI agent found it ran under a broad IAM role with keys far wider than the read-only access its job required. Agent frameworks inherit the developer's identity, making blast radius the union of everything that identity can reach. The practitioner community is being asked how it schedules agent identity audits and scopes permissions at creation.
An engineer auditing an internal AI agent found it ran under a broad IAM role with keys far wider than the read-only access its job required. Agent frameworks inherit the developer's identity, making blast radius the union of everything that identity can reach. The practitioner community is being asked how it schedules agent identity audits and scopes permissions at creation.
7 engagement1 sourceredditMon, Sep 21, 2026, 10:49 AM
Products·launched
Prompt to verify Supabase RLS policies via AI agents
A prompt that can be given to AI agents like Claude or Codex to automatically verify Row Level Security policies on every table in a Supabase project. It helps developers ensure their database security policies are correctly configured.
A prompt that can be given to AI agents like Claude or Codex to automatically verify Row Level Security policies on every table in a Supabase project. It helps developers ensure their database security policies are correctly configured.
1 engagement1 sourcehackernewsThu, Aug 13, 2026, 12:39 PM
Products·launched
Free SQL audit script to find RLS leaks in AI-built Supabase apps
A read-only SQL file that scans a Supabase database for five common Row Level Security misconfigurations introduced by AI coding tools like Lovable, Bolt, Cursor, and Replit. It helps developers and security auditors quickly identify data exposure risks without modifying any data.
A read-only SQL file that scans a Supabase database for five common Row Level Security misconfigurations introduced by AI coding tools like Lovable, Bolt, Cursor, and Replit. It helps developers and security auditors quickly identify data exposure risks without modifying any data.
1 engagement1 sourceredditFri, Jul 17, 2026, 09:29 PM
Releases·Tool Release
Developer launches pg-dry-run to preview Postgres writes from AI agents via xmin checks
A developer shared pg-dry-run on Hacker News, a tool that lets AI agents preview Postgres writes before committing them using xmin checks. It targets the practical problem of letting agents run safely against production databases. Interest so far is minimal, with low engagement on the launch post.
A developer shared pg-dry-run on Hacker News, a tool that lets AI agents preview Postgres writes before committing them using xmin checks. It targets the practical problem of letting agents run safely against production databases. Interest so far is minimal, with low engagement on the launch post.
2 engagement1 sourcehackernewsMon, Aug 31, 2026, 03:11 PM