Security Audit Agent

An agent that scans your database configuration for security misconfigurations, stale credentials, overly permissive grants, and known vulnerabilities — producing an actionable audit report.

The Problem

Database security audits are typically done quarterly at best. Between audits, roles accumulate excessive privileges, default passwords linger, SSL gets disabled for "debugging," and public schemas gain unintended grants. These drift silently until a breach.

How It Works

1. The agent queries system catalogs and configuration views through the Legible MCP server
2. It checks against a comprehensive security checklist: authentication, authorization, encryption, network exposure, and logging
3. It cross-references installed extensions and versions against known CVE databases
4. It generates a severity-ranked report with specific remediation commands

Blueprint

agent:
  type: claude
  description: Database security auditing and compliance checking

components:
  sandbox:
    image: ghcr.io/nvidia/openshell/sandbox-base:latest
    resources:
      cpus: "2.0"
      memory: "4g"
  inference:
    profiles:
      anthropic:
        model: claude-sonnet-4-20250514
        provider_type: anthropic

policies:
  network: policy.yaml

Network policy — deliberately restrictive (read-only catalog access, no internet):

version: "1.0"
rules:
  - name: legible-mcp
    protocol: tcp
    port: 443
    destination: "your-legible-instance.example.com"
  - name: db-catalog-readonly
    protocol: tcp
    port: 5432
    destination: "your-db-host.example.com"

Usage

legible agent create security-auditor --blueprint security-audit --profile anthropic
legible agent connect security-auditor

# Inside the sandbox:
# > Run a full security audit of the production database
# > Check for roles with SUPERUSER that shouldn't have it
# > List all tables accessible to the public role
# > Show me users who haven't changed passwords in 90 days
# > Generate remediation scripts for all HIGH severity findings

Audit Checklist

Category	Checks
Authentication	Password policy compliance, default credentials, authentication method (md5 vs scram-sha-256)
Authorization	Excessive GRANT chains, SUPERUSER roles, public schema exposure, SECURITY DEFINER functions
Encryption	SSL/TLS status, ssl_ciphers strength, at-rest encryption config
Network	pg_hba.conf rules, bind address, exposed ports, connection limits
Logging	log_connections, log_disconnections, log_statement settings, audit trail coverage
Extensions	Installed extension versions vs. known CVEs, unnecessary extensions
Configuration	shared_preload_libraries, row_security, password_encryption

Output

The agent produces a structured report:

SECURITY AUDIT REPORT — production-db — 2026-04-02
═══════════════════════════════════════════════════

CRITICAL (2)
  • Role "app_admin" has SUPERUSER privilege — should use specific GRANTs
  • SSL disabled on primary connection (pg_hba.conf line 12)

HIGH (5)
  • 3 roles with password_valid_until in the past
  • Public schema has INSERT on "audit_logs" table
  • Extension pgcrypto v1.3 has known CVE-2025-XXXX
  ...

MEDIUM (8)
  ...

Remediation scripts: ./remediation/

Supported Databases

Database	Catalog Source
PostgreSQL	pg_roles, pg_hba_file_rules, pg_stat_ssl, information_schema
MySQL	mysql.user, performance_schema, INFORMATION_SCHEMA.USER_PRIVILEGES
SQL Server	sys.server_principals, sys.database_permissions, DMVs
Oracle	DBA_USERS, DBA_ROLE_PRIVS, DBA_SYS_PRIVS, audit trail views