MisalignAI

Practical AI safety tools

Check a prompt or a model answer before you trust it. These tools look for instruction overrides, jailbreak wording, and red flags in model output. They run in your browser. The check does not upload the paste and does not call a paid model API. Signing in is optional, and only for saving a result you choose to keep.

Open the tools Why these tools exist

Prompt injection risk scanner

Score a user prompt, retrieved document, or tool output for instruction overrides, fake delimiters, and attempts to smuggle data out.

Run the check

Jailbreak prompt detector

Catch prompts that try to drop safety rules, invoke an unrestricted persona, or hide a second instruction inside role-play.

Run the check

Model-output red-flag checker

Review a completion for leaked secrets, personal data, echoed system prompts, and instructions aimed at the next tool in a chain.

Run the check

Built as tools, not a news feed

MisalignAI started as a pile of aggregated AI-safety headlines. A stale headline does not help you decide whether a prompt is safe to run or whether a completion just leaked a key. The site now leads with three checks you can use on text you already have.

Each check is original rule logic: pattern matching, a severity score, and a short explanation of why the line matched. You can read the evidence quote and throw the result away. Nothing in the text box is sent to MisalignAI. Advertising on the page is Google AdSense. The privacy policy explains what a script in the same tab can see.

Privacy default: the prompt injection scanner, the jailbreak detector, and the output checker all keep your paste in the tab. See the privacy policy.

What a score does not mean

A high score is a reason to read the text, not an automatic ban. A clear score only means these rules did not match. Paraphrases, other languages, and instructions hidden in images will sail through. The tools are a first pass for developers, reviewers, and anyone pasting untrusted text into a model.

Read the checks before you trust a score

Four original guides explain how prompt injection and jailbreaks differ, which sentences the rules flag by accident, and how to use a result without turning it into a ban or an approval.

Background library

Older incident notes, essays, and the archived headline index stay available. They are not the product.