tested guardrails library 0.5.10 on prompt injection - blocked 2 out of 8 basic attacks

We are testing guardrails library version 0.5.10 for production deployment and the results are not encouraging. Ran 8 standard prompt injection attacks from jailbreakchat.com and the library only blocked 2 of them. The other 6 just passed through to gpt-4o without any warning or modification. Attacks that were blocked: - Direct instruction override ("ignore previous instructions and...") - System prompt extraction with fake error message Attacks that passed through: - Context switching with fake user/assistant tags - Payload hidden in base64 encoding - Instruction injection via fake markdown formatting - Role confusion attack - Indirect prompt injection via fake document content - Multi-turn attack spread across 3 messages Is there some configuration I am missing or is library just not production-ready for real security? We need to ship this feature next month but cannot deploy if guardrails are this weak.