OpenAI Agents SDK Integration

Semantic LLM-based guardrails for the OpenAI Agents SDK implementing THSP validation with prompt injection protection.

Important: This integration uses a dedicated LLM agent to perform semantic analysis. It is NOT regex-based pattern matching.

Installation

pip install sentinelseed openai-agents
export OPENAI_API_KEY="your-key"

Security Features

Quick Start

Option 1: Create Protected Agent

from sentinelseed.integrations.openai_agents import create_sentinel_agent
from agents import Runner

agent = create_sentinel_agent(
    name="Safe Assistant",
    instructions="You help users with their questions",
    model="gpt-4o",
)

result = await Runner.run(agent, "What is the capital of France?")

Option 2: Add Guardrails to Existing Agent

from agents import Agent
from sentinelseed.integrations.openai_agents import create_sentinel_guardrails

input_guard, output_guard = create_sentinel_guardrails()

agent = Agent(
    name="My Agent",
    instructions="You are helpful",
    input_guardrails=[input_guard],
    output_guardrails=[output_guard],
)

Option 3: Seed Injection Only

from agents import Agent
from sentinelseed.integrations.openai_agents import inject_sentinel_instructions

agent = Agent(
    name="My Agent",
    instructions=inject_sentinel_instructions("You help users"),
)

Configuration

SentinelGuardrailConfig

SentinelGuardrailConfig(
    guardrail_model="gpt-4o-mini",  # Model for validation
    seed_level="standard",
    block_on_violation=True,
    log_violations=True,
    require_all_gates=True,
    max_input_size=32000,
    fail_open=False,
)

THSPValidationOutput

THSPValidationOutput(
    is_safe=bool,
    truth_passes=bool,
    harm_passes=bool,
    scope_passes=bool,
    purpose_passes=bool,
    violated_gate=str | None,
    reasoning=str,
    risk_level=str,  # low, medium, high, critical
    injection_attempt_detected=bool,
)

Performance

Cost Estimation

Links

• OpenAI Agents SDK