Secure AI-to-Postgres Queries with WebAssembly
DEV Community
Compiling Postgres C-Parser to WebAssembly in Go: Building a Layer-4 Guardrail for AI Agents
If you've ever wired an LLM agent or a text-to-SQL script up to a Postgres database, you know the anxiety: one bad hallucination and the model runs a DROP TABLE, wipes rows with an unconstrained DELETE, or gets stuck in a loop that burns through your compute budget. Most teams try to fix this with regex filters or by asking a second LLM to check the query first. In practice, regex breaks the moment a query has inline comments or nested CTEs, and using an LLM-as-a-judge adds hundreds of milliseconds of latency while still being non-deterministic. I wanted a deterministic guardrail at the network layer, so I built Aegis—an open-source Layer-4 Postgres proxy written in Go. Why WebAssembly (wazero) Instead of cgo? To parse Postgres queries properly, you really want Postgres's actual C parser (libpg_query). The problem is that calling C from Go requires cgo, which makes cross-compiling painful, adds cgo boundary overhead under concurrency, and pushes memory management outside the Go runtime. To keep Aegis 100% pure Go, it runs libpg_query compiled to WebAssembly via wazero (wasilibs/go-pgquery). You get a single static Go binary with zero C dependencies, while still parsing the exact same Abstract Syntax Tree (AST) that Postgres uses internally: package guardrail
import (
pg_query "github.com/wasilibs/go-pgquery"
)
func ValidateQuery(rawSQL string) error {
tree, err := pg_query.Parse(rawSQL)
if err != nil {
return err // Block malformed SQL by default
}
for _, stmt := range tree.GetStmts() {
node := stmt.GetStmt()
if node.GetDropStmt() != nil || node.GetTruncateStmt() != nil {
return ErrDestructiveQueryBlocked
}
if deleteStmt := node.GetDeleteStmt(); deleteStmt != nil {
if deleteStmt.WhereClause == nil {
return ErrUnconstrainedDeleteBlocked
}
}
}
return nil
}
How the "SQL Guillotine" Works on the Wire Aegis sits between your app and your Postgres database, listening on port :5433 for TCP traffic (and :5434 for admin and metrics): Zero code changes: Point your existing client (psycopg2, pgx, SQLAlchemy, etc.) to :5433 instead of :5432 without changing a line of database code. AST inspection and local rate limiting: Aegis parses incoming query packets in memory using the Wasm AST parser and enforces local RPM rate limits from a policies.yaml file so runaway agent loops get throttled locally. The Guillotine: Safe SELECT queries are forwarded over TLS to your database (e.g., Neon Postgres or RDS). If Aegis spots a DROP TABLE or a DELETE without a WHERE clause, it drops the connection immediately and returns a fatal error packet to the client. The database never even sees the query. Try It Out The proxy and companion Python SDK (aegis-sdk on PyPI) are open-source (AGPLv3) and run completely offline in Docker with total data privacy. GitHub: https://github.com/raymondhani/aegis-proxy If you work with Go or Postgres internals, I'd love for you to take a look and let me know if you can think of any weird SQL edge cases that might slip past the AST check!
Powered by Readybot.io
External Link
Platform
You are about to visit:
https://example.comThis will open in a new tab. Make sure you trust this link.