Revisiting the Analog Hole: Using OCR and other techniques to exfiltrate data

The talk asks whether optical character recognition can be used to pull data out of an environment. The speaker's core point is that any information a person must read has to be shown in an analog form, such as a monitor, printed page or projected slide, so a phone camera can capture it. Remote work makes this harder to spot, and a company laptop may leave no log of the capture.

The walkthrough covers how OCR works: converting the image to black and white, correcting rotation and scale, segmenting characters, words or paragraphs, and then recognizing them with older template matching or newer neural models. Photos taken at an angle need perspective correction first. Cloud and office tools were tested on captured images. Clean, high-contrast text came through well, while inverted text, background noise and skew caused errors. Other routes covered include screenshots, HDMI capture, and accessibility interfaces that dump application text directly. Encoded data such as base64 is hard for OCR engines to read reliably.

For defenders, the speaker suggests showing each user only the data they need, giving different users slightly different content so a leak can be traced, using visible but unobtrusive watermarks, and deploying endpoint data-loss tools that keep activity logs. Watermarks can be lost once text is re-extracted, so cryptographically verifiable data is a sturdier control than anything visual. The closing takeaway is that OCR-based exfiltration is most useful for mid-sized leaks or when a remote desktop is the only access path.