How do I use InjectShield with the OpenAI Agents SDK?

The OpenAI Agents SDK (2025+) supports input_guardrails and output_guardrails natively — InjectShield ships as a drop-in guardrail. Install pip install injectshield-openai-agents, then:

from agents import Agent, input_guardrail, output_guardrail
from injectshield_openai_agents import InjectShieldInputGuardrail, InjectShieldOutputGuardrail

agent = Agent( name="research_agent", instructions=SYSTEM_PROMPT, tools=[web_search, fetch_url], input_guardrails=[InjectShieldInputGuardrail()], output_guardrails=[InjectShieldOutputGuardrail()], ) ```

The input guardrail scans the user message before the model runs; the output guardrail scans the model's response before any tool call is dispatched. Both raise GuardrailTripwireTriggered on a positive verdict, which the SDK handles by halting the run and surfacing the trip reason to the caller — exactly the right behavior for production.

For function-calling tool outputs (the highest-risk surface in agent stacks per OWASP LLM01 + LLM07), wrap each tool function with @injectshield_tool — the wrapper classifies the tool's return value with context: "tool_output" before it re-enters the model's context. Combine with the SDK's tool_use_behavior allowlist to bound what the agent can do after a positive verdict.

Logs flow to OpenAI's tracing dashboard and (with INJECTSHIELD_DASHBOARD=true) to injectshield.dev/dashboard for cross-stack monitoring. Reference implementation at injectshield.dev/docs/openai-agents-sdk.