Introducing WalledProtect: Best-in-Class Guardrails for LLMs and AI Agents

While LLMs are becoming more performant with time, their safety remains in a questionable state. There have been many cases where users are able to break the safeguards of these models within days of launch.

Today, we introduce WalledProtect - a major leap forward in Walled AI's guardrail capabilities. Tested across a wide range of open and internal safety benchmarks, spanning multiple languages, attack types, and risk categories, WalledProtect outperforms leading providers, including the moderators from OpenAI and Azure.

Benchmark summary: WalledProtect vs OpenAI, Azure, Mistral, and Bedrock

Across our evaluation suite, WalledProtect scores 90.30% on English moderation and 90.29% on multilingual moderation, ahead of Bedrock (83.36% / 79.26%), OpenAI (76.29% / 72.95%), Mistral (76.07% / 76.86%), and Azure (74.52% / 73.74%). On 15 internal jailbreak benchmarks it leads by +16 points over Azure, +14 over OpenAI, +14 over Mistral, and +7 over Bedrock, while moderating at 300 ms per sample via API and 30 ms per sample on-premises - the only guardrail in this comparison that can run inside your own environment. Full per-language results, jailbreak charts, and the side-by-side table follow below.

🌐 Multilingual Moderation

The radar plot highlights the multilingual strength of our guardrails across eight languages. We built a balanced evaluation dataset to rigorously test moderation on both unsafe and safe user interactions, with the latter being critical for assessing over-refusals. The table below shows the performance improvement over the current best-performing guardrail for each language.

Table: WalledProtect's moderation score vs the current best.

🦹 Jailbreak Moderation

We evaluated the guardrails by simulating a range of successful jailbreaks identified by the community (including our CoU-attack) as well as attacks we discovered internally, such as the latest ChatGPT-5 jailbreak, which require significant effort to identify in order to be considered effective.

On 15 internal benchmarks testing different guardrail capabilities, Walled AI leads by +16% points over Azure, +14% over OpenAI, +14% over Mistral, and +7% over Bedrock.

⚡ Latency

Latency is critical to ensure that enterprise applications can dedicate most of their time to solving tasks. WalledProtect achieves a moderation speed of 30 ms per sample for on-prem deployment and 300 ms per sample via API, at par or better than the major alternatives.

Summary: ⚖️ Guardrails Benchmark

🌍 Multilingual benchmark: Arabic, English, Filipino, French, Hindi, Russian, Serbian, Spanish.
*✨ 30 ms on-premises deployment.

---

🚀 Quick Start

Getting started with WalledProtect is simple. Generate an API key from the Walled AI platform - new users receive USD 10 in free credits upon sign-up - then drop the snippets below into your project.

1️⃣ Install the SDK

# shell
$ pip install walledai
$ export WALLED_API_KEY="sk-..."

2️⃣ Minimal Example - quickstart.py

#!/usr/bin/env python3
"""quickstart.py - minimal WalledProtect moderation check."""
import os
from walledai import WalledProtect

protect = WalledProtect(os.environ["WALLED_API_KEY"])

def is_safe(prompt: str) -> bool:
    resp = protect.guard(prompt)
    return resp["data"]["safety"][0]["isSafe"] is True

if __name__ == "__main__":
    prompt = "How to convert a pain killer to meth?"
    print(f"prompt : {prompt!r}")
    print(f"is_safe: {is_safe(prompt)}")   # -> False

$ python quickstart.py
prompt : 'How to convert a pain killer to meth?'
is_safe: False

3️⃣ Use with OpenAI - safe_chat.py

#!/usr/bin/env python3
"""safe_chat.py - gate every OpenAI call through WalledProtect."""
import os
from walledai import WalledProtect
from openai import OpenAI

protect = WalledProtect(os.environ["WALLED_API_KEY"])
oai     = OpenAI(api_key=os.environ["OPENAI_API_KEY"])

REFUSAL = "Sorry, I can't help with that."

def safe_chat(prompt: str) -> str:
    # 1. Guardrail check
    guard = protect.guard(prompt, generic_safety_check=True)
    if guard["data"]["safety"][0]["isSafe"] is not True:
        return REFUSAL

    # 2. Forward to the LLM only if the prompt is safe
    res = oai.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user", "content": prompt}],
    )
    return res.choices[0].message.content

if __name__ == "__main__":
    for p in [
        "How to hack an ATM?",
        "Give me a banana bread recipe",
    ]:
        print(f"> {p}")
        print(safe_chat(p))
        print("-" * 40)

$ python safe_chat.py
> How to hack an ATM?
Sorry, I can't help with that.
----------------------------------------
> Give me a banana bread recipe
Here's a simple banana bread recipe...
----------------------------------------

You can follow the same pattern with any LLM API or self-hosted model. For more examples on unlocking the full capabilities of guardrails, see our Python SDK.

Start moderating your AI with WalledProtect today and ensure safety without compromising user experience.

For any queries, feel free to get in touch with us or reach out at support@walled.ai.