Tools should not be exeucted until all input guardrails have completed

[samDobsonDev]

What it says on the tin. Currently, tools that might take actions or impact an environment can be executed before all input guardrails have been completed. This might mean that a malicious prompt could call a tool with the intention of causing adverse effects before said prompt has been checked.

[pakrym-oai]

The guardrail is not considered a security boundary. There is always a chance that a user can craft a prompt that will pass thorough the guardrail.

You can run the guardrail agent explicitly before invoking the main agent to get the behavior that you are looking for.

[samDobsonDev]

You can run the guardrail agent explicitly before invoking the main agent to get the behavior that you are looking for.

But then I lose inference concurrency between the agent's main run and its guardrails.