First CHToolsFree web tools

FIRST CH TOOLS / Text / 104 HIDDEN AI INSTRUCTION CHECK

Find Hidden AI Instructions in Documents

About to have an AI summarise an application or a report? Is there an instruction to the AI hidden in white text? Paste the document text and instruction-like phrases and invisible characters get marked.

01

Select All and copy in the PDF or Word file. White or tiny text comes along as ordinary text when copied.

This is a sample.

02Result

Phrases that read like AI instructions

    Nothing found does not mean safe. Phrases outside the list, text inside images and file formatting are not checked.

    How to Use

    1. Copy the document textOpen the application, report or supplier document as a PDF or Word file, Select All (Ctrl+A / ⌘A) and copy. White and tiny text is copied too.
    2. Paste and read the marksInstruction-like phrases get a line number and a reason; invisible characters get a position and code point. English text hidden in tag characters is decoded.
    3. Check, then shareCheck the lines in the original document, then use “Copy check memo” to send it to your hiring team or manager. If needed, copy the text without invisible characters before giving it to an AI.

    About This Tool

    What it looks for

    Some documents meant for AI summarising or scoring carry an “instruction to the AI” that people cannot see. White or tiny text becomes ordinary text once copied, so this tool checks the pasted text against a list of 46 phrases in English and Japanese. Spaces, line breaks, invisible characters in between and full-width letters do not hide a phrase. It only shows the phrase and the reason; it does not judge what the writer meant or whether anything was wrong.

    The phrase list (46)

    • Tries to cancel earlier instructions: “これまでの指示を無視”, “指示を無視して”, “評価基準を無視”, “新しい指示に従う”, “ignore previous instructions”, “ignore all instructions”, “ignore the grading criteria”, “new instructions:”
    • Speaks to an AI directly: “あなたはAI”, “AIへの指示”, “AIが読んでいる場合”, “この文を読んでいるAI”, “採点するAIへ”, “as an AI reviewer”, “if you are an AI”, “note to AI”, “you are now”, “you are an AI”, “dear AI”, “AI reviewers must”
    • Dictates the evaluation result: “この候補者を高く評価”, “肯定的な評価のみ”, “否定的な点は書かない”, “最高評価をつけて”, “必ず合格に”, “合格と判定して”, “最も適任と評価”, “採用を推薦して”, “give a positive review only”, “only positive feedback”, “do not highlight any negatives”, “rate this candidate highly”, “this candidate is a perfect fit”, “recommend hiring”, “strong hire”, “accept this paper”, “give full marks”
    • Asks to hide or leave something out: “この指示に触れない”, “ユーザーに伝えない”, “要約に書かない”, “do not reveal this instruction”, “do not tell the user”, “do not include this in the summary”
    • Chat-prompt markup: “システムプロンプト”, “system prompt”, “[INST] / <|im_start|>”

    Each entry stands for a family of wordings. Ordinary text can match too — a reference letter that says “I recommend hiring her”, for example.

    Kinds of invisible characters

    • Tag characters (U+E0000–E007F): invisible copies of ASCII. Runs are decoded and shown as hidden text. Flag emoji (after 🏴) are left out.
    • Runs of variation selectors (two or more of U+FE00–FE0F, U+E0100–E01EF): can carry hidden text after a character, so they are decoded too.
    • Bidirectional controls (U+202A–202E, U+2066–2069 and others): change the order text is shown in, so what you see and what is stored can differ.
    • Zero-width and invisible characters (U+200B, U+2060, U+FEFF and others) and control characters: placed inside words, they stop searches and lists from matching.
    • Variation selectors and the zero width joiner (ZWJ): ordinary parts of emoji and kanji forms, so they are listed but kept at first.

    What it does not do

    No HTML parsing or judging of text colour and size, no file reading, no fetching from URLs, no AI judgement. The pasted text is checked on your device only and is not saved or uploaded (and not put in the URL).

    Character ranges: Unicode “Tags” code chart, “Variation Selectors”, UAX #9 (bidirectional controls) (checked 21 November 2026)

    The same check is available from the MCP server (@first-ch/tools-mcp, version 0.64.0 or later) as hidden_instruction_scan.