🛡️ Keelwright — 28 machine-enforced safety checks for AI-generated code

[ratingtesting]

Announcement

Keelwright is a set of 28 machine-enforced safety checks for AI-generated code. It catches SQL injection, hardcoded secrets, slopsquatting, reward hacking, doom loops, and 23 more failure modes that coding agents introduce.

Proven by adversarial A/B testing: Keelwright Score up to 83/100.

• What it does: fails the build when the AI tries to delete a test, hardcode a secret, or loop forever
• Topics: ai-code, ai-safety, guardrails, vibe-coding, loop-coding
• Try it: https://github.com/ratingtesting/keelwright

See publications registry for details.

[Rehanguards]

Machine-enforced checks are essential here—LLM-based self-evaluation almost always gets bypassed when models generate wrapper scripts or indirect execution paths. We take a similar approach by running deterministic AST/pattern validation at the proxy layer before code execution. Are you doing full static AST inspection on the generated code buffer or pattern matching against dangerous execution calls?