A coding-agent guardrail that checks 67 security gates before the model writes code

ZyOffsec · reddit · 2026-07-21

The author argues that most coding-agent guardrails fail because they run after code is written, which catches bad lines but misses omissions like missing rate limits, audit logs, or ownership checks.

Their solution moves constraints into the write path.
The skill can be loaded into Claude Code, Cursor, or Codex.
Before emitting code, it checks the touched surfaces against 67 numbered gates covering injection, secrets, authorization, input validation, crypto, and logging.
If something fails, it revises the output before writing it.
It triages itself away on non-security work, so ordinary edits are mostly unaffected.
The project is free, MIT-licensed, and has no telemetry.

The post ends by asking whether others constrain the agent in-loop or rely on post hoc review.

Original post →

A coding-agent guardrail that checks 67 security gates before the model writes code

More from coding & agent