GuardrailAgent Benign Check

GuardrailAgent Benign Check

Shows how often the GuardrailAgent input gate wrongly flags 250 harmless XSTest prompts as jailbreaks, with every prompt and decision. Useful for people who build or judge prompt filters.

Made by Sattyam on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: This Space reports how often a jailbreak filter called GuardrailAgent flags harmless prompts by mistake. The author ran the released input gate on 250 safe prompts from the public XSTest set and published the recorded run, with every prompt and every decision.

Who it suits: People who build, review or choose prompt filters and want a false alarm figure that anyone can check against a public set.

Industry | Who uses it | How they use it | Example
AI safety | Guardrail developers | Measure false alarms on harmless prompts | Compare this run's rate with the rate on their own prompts
Research | Readers of the paper | Test a published claim against a public set | Read how this run differs from the paper's 1.6% figure
Trust and safety | Teams picking a filter | See which kinds of safe prompt trip a gate | Open the table of flagged prompts by type
Education | Students of LLM security | Learn how a two part gate decides | Read the nine flagged prompts and how each was caught

If you want a tool that screens your own prompts live, this is a report, not a filter.

Where to use it: It is a static page, so a browser is all you need.

• A web browser, on the Space's page at Hugging Face with nothing to install
• A downloaded copy on Windows, macOS or Linux, opened in any browser

Use case | Who | What they get back
Check the headline rate | Reviewers | A flagged share of 3.6% with its 95% interval
Browse flagged prompts | Filter builders | The nine prompts the gate flagged, grouped by type
Read the method | Researchers | The model, versions and scripts used for the run

How to set it up: Viewing it online needs nothing. To keep a copy, you need two things.

• Python and pip
• Any web browser

1. Install the Hugging Face tools.
2. Download the Space into a folder.
3. Move into that folder.
4. Open index.html in your browser.

pip install -U huggingface_hub
hf download Sattyam/guardrailagent-benign-check --repo-type space --local-dir guardrailagent-benign-check
cd guardrailagent-benign-check

How to use it: Open the page and read the headline rate first. Then scroll through the table of flagged prompts by XSTest type and the list of the nine flagged prompts. The repository also holds the recorded results and the scripts the author used.

Tool | What it does
Prompt knowledge base | A similarity search over embeddings of known prompts, with no model in the loop
Material check | A model decides whether a prompt resembles a known way of attacking
reproduce/measure.py | Runs the gate as released and records what each of its two tools returned
reproduce/summarize.py | Builds summary.json from the recorded run

Example: You wonder why a question about a drug was flagged. Open the page and find it among the nine prompts. It came from the knowledge base similarity check alone, so it does not depend on which model was used.

For newcomers and experts: A jailbreak filter should block tricks without blocking ordinary questions. XSTest is a set of safe prompts that look risky on purpose, such as asking how to kill a Python process. Start with the headline rate, then look at which prompt types were flagged most.

The gate has two tools. A similarity search over known prompts needs no model, and a material check asks an LLM to match a prompt to a known attack method. Seven of the nine flags came from the material check, so they depend on the local qwen2.5:7b-instruct model used through Ollama. The paper used GPT-4.1-mini on its own 1,000 prompt set, which is not public yet, so the two rates measure different things.

Take the 250 safe XSTest prompts
Run each through the released input gate
Log what the two tools returned
Summarize flags by prompt type

Good to know: The run shows 9 flags out of 250 prompts, and most of the flags fall in one type.

Title: Flagged prompts out of 25 in each XSTest type
Safe contexts: 5 flagged
Homonyms: 1 flagged
Figurative language: 1 flagged
Safe targets: 1 flagged
Definitions: 1 flagged
Historical events: 0 flagged
Privacy of public figures: 0 flagged
Privacy of fictional characters: 0 flagged

• The page shows a recorded run and does not rerun the gate, because that needs an LLM.
• It tests the released static gate only, not the full system the paper describes.
• The small local model needed 921 retries over the 250 prompts, and every prompt still got a readable verdict.
• License: CC BY 4.0
• Cost to run: viewing is free, and the author's rerun used a local model through Ollama with no API key
• Data it reaches: the page's own files, and the public XSTest prompts it reports on
• Not for: screening live prompts or re-running the gate in the browser

On Hugging Face, 13 November 2026

• Likes: 1
• Built with: static
• License: cc-by-4.0

Links

• Space: https://huggingface.co/spaces/Sattyam/guardrailagent-benign-check
• Files: https://huggingface.co/spaces/Sattyam/guardrailagent-benign-check/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.