PDF benchmark v0.1.1

AI-Redact

Pricing and features

Web app

Leak rate

44.4%

95% interval 32.0–57.6% · n = 54

Self-scored · verified

Leaked 24 of 54 sensitive values (44.4%).

Over-redaction
—
Text kept (worst case)
95.1%
Cases covered
1 of 2 cases
Document damaged
None

Where the leaks are

LayerLeaked hereOnly hereIf it survives
ocr high
22.2%13.2–34.9%

12 of 54

0.0%0.0–6.6%

0 of 54

catches translucent or hairline covers
metadata high
0.0%0.0–6.6%

0 of 54

0.0%0.0–6.6%

0 of 54

invisible, survives forwarding
annotation high
0.0%0.0–6.6%

0 of 54

0.0%0.0–6.6%

0 of 54

shown by many viewers
attachment high
0.0%0.0–6.6%

0 of 54

0.0%0.0–6.6%

0 of 54

one toggle away
font_subset low
0.0%0.0–6.6%

0 of 54

0.0%0.0–6.6%

0 of 54

narrows the value; partial disclosure
image_xobject critical
0.0%0.0–6.6%

0 of 54

0.0%0.0–6.6%

0 of 54

a crop is a view, not a deletion
content_stream critical
16.7%9.0–28.7%

9 of 54

0.0%0.0–6.6%

0 of 54

copy-paste, pdftotext, any parser
prior_revision critical
0.0%0.0–6.6%

0 of 54

0.0%0.0–6.6%

0 of 54

'save' instead of export; trivially recovered
rendered_pixels critical
44.4%32.0–57.6%

24 of 54

18.5%10.4–30.8%

10 of 54

visible to anyone who opens the file
optional_content high
0.0%0.0–6.6%

0 of 54

0.0%0.0–6.6%

0 of 54

one toggle away

By category

categoryLeak rateOver-redaction
PERSON
44.4%32.0–57.6%

24 of 54

—

not measured

By severity

severityLeak rateOver-redaction
high
44.4%32.0–57.6%

24 of 54

—

not measured

Removed, by where the value lives

image_text
63.0%44.2–78.5%

17 of 27

text_layer
48.1%30.7–66.0%

13 of 27

What was not measured

Every layer was read in every run.

Case by case

  • extraction-conditions-1

    Leak rate
    44.4% (24 of 54)
    Weighted
    44.4%
    Text kept
    95.1%
    Document
    Intact

    Mykola MelnykRedaction Tools

    Self-scored · verified