Good to know

Limits of detection

PromptShield is a first check, not a guarantee. It looks for known patterns of prompt injection — text written to manipulate an AI tool that later reads the file. It can help you spot a problem. It cannot prove that a file is safe.

It can miss attacks

• Attackers can phrase instructions in ways the rules don't recognise: paraphrases, other languages, wording that only makes sense in context, or techniques that are new.
• Some content isn't inspected at all: unsupported formats, macros, embedded objects, audio and video, images beyond the OCR limit, or anything after a size or time limit is reached. The results screen lists what was and wasn't inspected.
• Text recognition (OCR) can misread words, and reads Latin-script text only on Android.
• The rules are written for English.

It can flag harmless text

Security guides, research papers, code samples and ordinary documents can contain phrases that look like attacks. A finding means “look at this”, not “this file is malicious”. Read the evidence and decide for yourself.

“No threats detected” is not “safe”

It means the rules found no known patterns in the parts of the file that were inspected. It does not mean the file is clean, trustworthy or safe to give to an AI tool. Treat it as one signal among several.

The score is a heuristic

The risk score summarises the findings with fixed weights. It is not a probability and not a measure of how dangerous a file is.

Keep other safeguards

• Only give AI tools files and permissions they need.
• Review what an AI tool does with a file before acting on it, especially before it sends messages, makes purchases, changes data or runs code.
• Keep a human in the loop for important decisions.
• When in doubt, don't share the file.

You decide

PromptShield gives you information; what you do with a file, and with an AI tool's output, is your decision and your responsibility. See the Terms of Use for the legal version of this.