Faint ink
Low-contrast text that is easy to miss. Local contrast amplification and English OCR can reveal it; detection depends on the image and can produce false positives.
Prompt-injection forensics for images
Reveal the text an AI can read but you cannot see.
Images can carry instruction-like text in faint ink, transparent colour, simple low-bit messages or metadata. What reaches an AI depends on the model and its preprocessing pipeline. Inspect these channels in your browser with local English OCR and explainable checks.
Drop an image, paste one, or choose a file
PNG, JPEG, WebP, GIF or AVIF, when supported by your browser. Supported PNGs preserve raw transparency; exact low-bit checks require 8-bit PNG. Up to 60 MiB and 40 megapixels.
Five places instructions hide
Under transparency
Colour stored in fully transparent pixels. Pipelines that drop the alpha channel suddenly “see” it.
Low-bit messages
Probe four simple bit layouts at the start of an 8-bit PNG for readable ASCII messages. Models do not automatically decode these messages.
Metadata
Supported PNG text chunks, selected EXIF text tags, XMP values and JPEG or GIF comments that some pipelines may pass to a model. Other fields and formats can be missed.
Instruction language
Detected text is checked against English rules for override attempts, concealment, data exfiltration, actions and rating manipulation. Scores are review cues, not attack probabilities.