Subtext started where most inspection stops: documents, images, and archives. Those are the files people think of when they think about data leaving an organization. But secrets don’t only travel as text on a page. A password read aloud in a voice memo is exfil. A credential block hidden inside a media file is exfil. And an AI transcribing that audio will happily surface every word of it.

So media is now a first-class citizen in the pipeline. When Subtext meets an audio or video file, it does the same thing it does with everything else. It reduces the file to the text it actually carries, and then runs that text through the same content checks as a document.

What it looks at

  • The text already inside the container. Media files carry metadata, tags, embedded titles, and cover art. Any content sitting in those layers, including text buried in an embedded image, is extracted and inspected. This runs on every media file, always.
  • What is spoken. Speech in the audio is transcribed and read. A secret said out loud is now something a boundary can see, not just something a person would have had to sit and listen for.
  • What is on screen. Frames are pulled from the video and read as images, so a credential on a shared screen, a document held up to a camera, or text that only ever appears in the picture is inspected the same way text in a document is.
  • What is drawn into the sound. Text can be painted into a spectrogram, readable when you look at the audio as an image and invisible when you listen to it. Subtext renders that view and reads it through the same path it uses for any other image.

Honest about the envelope

A capability like this is only trustworthy if it’s honest about its own edges. When a piece of audio can’t be transcribed cleanly, Subtext says so. It declares what it was and wasn’t able to read, rather than returning a confident “nothing here” that quietly means “I couldn’t look.” A verdict is only as honest as what it actually inspected, and that applies to what it couldn’t inspect just as much.

The point isn’t that audio is a common exfil channel today. It’s that once an AI reduces every modality to text, the modality stops being a hiding place, but only if the thing guarding the boundary reduces it the same way. Now it does.