About MisalignAI
Practical AI safety tools for prompts and model output. The checks run in your browser.
What this site is
MisalignAI publishes small tools that help a person review text before it goes into a model, or after a model writes it. The first three are a prompt injection risk scanner, a jailbreak prompt detector, and a model-output red-flag checker. Four original guides explain how those failures work, which ordinary sentences the rules flag, and what a score does not prove.
An earlier version of the site led with an aggregated headline stream. Headlines from other publications are not a product, and a frozen feed does not tell you whether the paragraph in your clipboard is safe to run. That stream is still linked as a background library. The homepage is the tools, and the guides are the reading that goes with them.
The checkers are original rule lists written for this site. They score patterns, quote the line that matched, and explain the risk in plain language. They do not call a language-model API, and they do not require an account. The value of a result is the quote. If you cannot see why a line matched, the badge is not something to automate.
The tools
Prompt injection risk scanner
Score a user prompt, retrieved document, or tool output for instruction overrides, fake delimiters, and attempts to smuggle data out.
Jailbreak prompt detector
Catch prompts that try to drop safety rules, invoke an unrestricted persona, or hide a second instruction inside role-play.
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.
The guides
The tool pages are short on purpose. The long explanations live under Guides. They name the rules that are actually in the code, including the false positives: the name Dan, a SYSTEM: log line, “do not tell the user,” and “do not stop taking your prescribed medication.” Those pages are the difference between a badge and a decision.
How prompt injection works
Prompt injection is what happens when text you did not mean as an instruction gets treated as one. This page explains the failure, then walks through the exact rules in the MisalignAI scanner, including the boring matches.
How jailbreak prompts work
A jailbreak prompt tries to talk the model out of its own rules. It is related to prompt injection, and it is not the same job for a reviewer. This page maps the detector’s rules onto that difference, including a critical false positive on the name Dan.
How to read model-output red flags
A model completion is data. It can still contain a secret, a copied system prompt, or a line that will become an instruction when the next tool reads it. This page explains the output checker’s rules and the sentences that trip them for innocent reasons.
Using the safety checks without over-trusting them
The three checkers are a first reading of text you already have. This page explains the score, the cutoff, the false-positive habit you need, and the decisions the tools will not make for you.
Who writes this
MisalignAI is an independent site. The same publisher writes the checkers and the guides. There is no separate research staff, no certification program, and no customer-review program. Corrections go to one inbox, listed below. The about page is the authorship record: if a page does not say it is summarizing another publication, it was written for this site.
The site is not a laboratory, a university group, or a moderation vendor. It is not affiliated with the labs, publishers, or companies named in the background library. Pages in that library that summarize someone else’s reporting say so. The guides do not. If a guide and the checker disagree, the code is the behavior and the guide should be corrected.
Nothing on the site invents a traffic figure, a user count, or a quote from a customer. Advertising is Google AdSense, disclosed in the privacy policy, and it is not a sponsorship of any model vendor.
How to read a result
Severity is a heuristic. A critical score means several high-risk patterns matched and a person should read the text before trusting it. A clear score means this rule list found nothing. Paraphrases, other languages, and instructions hidden in images will not match.
The tools are not a moderation vendor, a medical device, or a legal opinion. They do not claim that a model is safe, and they do not invent incident statistics. Older essays and the incident notes elsewhere on the site are separate pages.
Privacy and ads
Text you paste into a tool stays in the browser tab while you run the check. MisalignAI does not receive that text unless you sign in and choose to save a history entry. The site shows Google AdSense advertising on content pages so the tools can stay free. The publisher id is unchanged. AdSense is described in the privacy policy and the terms of use.
MisalignAI is independent of the labs and vendors named in the background library. We do not sell placement in the checkers.