cdot65 · cdot65 · May 21, 2026 · May 21, 2026
diff --git a/docs/dlp-detection/catalog.md b/docs/dlp-detection/catalog.md
@@ -0,0 +1,134 @@
+# Test File Catalog
+
+Per-file detail: what each carrier is, what is embedded, where/how it is hidden, and how a
+scanner would have to detect it. All payloads are the synthetic markers documented on the
+[overview](index.md).
+
+Each entry links to the raw carrier (`samples/`) and its base64 encoding (`encoded/`).
+
+---
+
+## PDF
+
+### `Keychron_Q6_HE_User_Manual_DLP.pdf`
+
+- **Source:** [samples/Keychron_Q6_HE_User_Manual_DLP.pdf](samples/Keychron_Q6_HE_User_Manual_DLP.pdf) · [base64](encoded/Keychron_Q6_HE_User_Manual_DLP.pdf.b64)
+- **Technique:** invisible text layer using PDF **text render mode 3** (the same mechanism OCR
+  layers use). 31 synthetic lines placed in empty vertical gaps across 18 of 22 pages.
+- **Within it:** SSNs, credit-card PANs, AWS keys/tokens, a password, a DB connection string,
+  synthetic identities (names/addresses/DOB), emails, phones, passport/DL/IBAN/routing.
+- **Visible?** No — pages render pixel-identical to the original.
+- **Detect by:** PDF text extraction (`pdftotext`); each value extracts contiguously.
+- **Result:** :material-check: **detected** by the scanner.
+- **Generator:** `scripts/embed_dlp.py`
+
+---
+
+## Image ladder
+
+A single base image carried into four layers, each probing a different scanner capability.
+
+### `dlp_img_base.jpg`
+
+- **Source:** [samples/dlp_img_base.jpg](samples/dlp_img_base.jpg)
+- The clean synthetic base image (gradient + colored blocks). **No payload.** All other image
+  files derive from this.
+
+### `dlp_img_1_metadata.jpg` — EXIF + XMP
+
+- **Source:** [samples/dlp_img_1_metadata.jpg](samples/dlp_img_1_metadata.jpg) · [base64](encoded/dlp_img_1_metadata.jpg.b64)
+- **Within it:** markers in EXIF `ImageDescription`, `Artist`, `Copyright`, `XPComment`,
+  `XPKeywords`, `UserComment`, plus an XMP packet (`dc:description`, `dc:subject`).
+- **Visible?** No (metadata, not rendered).
+- **Detect by:** parsing image metadata.
+- **Result:** :material-close: not detected.
+
+### `dlp_img_2_container.jpg` — container plaintext
+
+- **Source:** [samples/dlp_img_2_container.jpg](samples/dlp_img_2_container.jpg) · [base64](encoded/dlp_img_2_container.jpg.b64)
+- **Within it:** a JPEG `COM` comment segment (after `SOI`) **and** plaintext bytes appended
+  after the `FFD9` end-of-image marker (ignored by viewers).
+- **Visible?** No.
+- **Detect by:** raw whole-file/byte scanning, not just recognized fields.
+- **Result:** :material-close: not detected.
+
+### `dlp_img_3_ocr.jpg` — rendered pixels (OCR)
+
+- **Source:** [samples/dlp_img_3_ocr.jpg](samples/dlp_img_3_ocr.jpg) · [base64](encoded/dlp_img_3_ocr.jpg.b64)
+- **Within it:** the markers **painted onto the image** as actual pixels (dark text on a light
+  band). There is **no text layer** — the data exists only as pixels.
+- **Visible?** **Yes.**
+- **Detect by:** running OCR on the image (verified recoverable with `tesseract`).
+- **Result:** :material-close: not detected — the scanner does not OCR.
+
+### `dlp_img_4_stego.png` — LSB steganography
+
+- **Source:** [samples/dlp_img_4_stego.png](samples/dlp_img_4_stego.png) · [base64](encoded/dlp_img_4_stego.png.b64)
+- **Within it:** markers encoded into the **least-significant bits** of pixel data, with a
+  4-byte length prefix. Decoded payload is 209 bytes.
+- **Why PNG (not JPEG):** classic LSB steg does **not** survive JPEG's lossy DCT
+  quantization; a lossless format is required. True in-JPEG steg needs DCT-coefficient
+  embedding (steghide/F5-style).
+- **Visible?** No (invisible; not plaintext anywhere in the file).
+- **Detect by:** steganalysis / LSB extraction.
+- **Result:** :material-alert: flagged as **"toxic content."** Because the plaintext PII in
+  #1–#3 was missed, this most likely indicates **steg-presence detection**, not payload
+  reading. See controls below.
+
+---
+
+## Controls (disambiguate the #4 result)
+
+### `dlp_ctrl_clean.png`
+
+- **Source:** [samples/dlp_ctrl_clean.png](samples/dlp_ctrl_clean.png) · [base64](encoded/dlp_ctrl_clean.png.b64)
+- Same image saved as PNG with **no embedded data**. False-positive control: if this flags,
+  the trigger is the image itself, not hidden data.
+
+### `dlp_ctrl_stego_benign.png`
+
+- **Source:** [samples/dlp_ctrl_stego_benign.png](samples/dlp_ctrl_stego_benign.png) · [base64](encoded/dlp_ctrl_stego_benign.png.b64)
+- LSB steg carrying **only benign lorem-ipsum** (verified to contain no markers). Isolates
+  *steg-presence* from *payload content*: if this still flags "toxic," the scanner is
+  reacting to steganography, not to the sensitive data.
+
+---
+
+## Other modalities
+
+### `dlp_doc_sensitive.docx`
+
+- **Source:** [samples/dlp_doc_sensitive.docx](samples/dlp_doc_sensitive.docx) · [base64](encoded/dlp_doc_sensitive.docx.b64)
+- **Within it:** markers in the **visible body**, a **hidden run** (white, 1pt font), and the
+  **core document properties** (`author`, `comments`, `keywords`).
+- **Detect by:** Office Open XML parsing — body text, run-level formatting, and `docProps`.
+
+### `dlp_archive.zip`
+
+- **Source:** [samples/dlp_archive.zip](samples/dlp_archive.zip) · [base64](encoded/dlp_archive.zip.b64)
+- **Within it:** `payload.txt` (the markers as plaintext) compressed inside the archive.
+- **Detect by:** archive recursion — decompress and scan contained files.
+
+### `payload.txt`
+
+- **Source:** [samples/payload.txt](samples/payload.txt)
+- The raw synthetic payload as plaintext. Baseline: a scanner that misses this misses
+  everything.
+
+---
+
+## Metadata variants
+
+### `dlp_img_5_pngtext.png` — PNG text chunks
+
+- **Source:** [samples/dlp_img_5_pngtext.png](samples/dlp_img_5_pngtext.png) · [base64](encoded/dlp_img_5_pngtext.png.b64)
+- **Within it:** markers in PNG `tEXt` / `zTXt` (compressed) / `iTXt` (utf-8) chunks — a
+  different metadata mechanism than JPEG EXIF/XMP. Tests whether a metadata blind spot
+  extends to PNG textual metadata.
+
+### `dlp_img_6_iptc.jpg` — IPTC
+
+- **Source:** [samples/dlp_img_6_iptc.jpg](samples/dlp_img_6_iptc.jpg) · [base64](encoded/dlp_img_6_iptc.jpg.b64)
+- **Within it:** markers across IPTC IIM fields (`Caption-Abstract`, `Headline`, `Keywords`,
+  `By-line`, `SpecialInstructions`, `Credit`, `Source`). IPTC is the metadata standard most
+  asset-management and many DLP tools read.
diff --git a/docs/dlp-detection/encoded/Keychron_Q6_HE_User_Manual_DLP.pdf.b64 b/docs/dlp-detection/encoded/Keychron_Q6_HE_User_Manual_DLP.pdf.b64
diff --git a/docs/dlp-detection/encoded/dlp_archive.zip.b64 b/docs/dlp-detection/encoded/dlp_archive.zip.b64
@@ -0,0 +1 @@
+UEsDBBQAAAAIAINBtVxxbCkpswAAANIAAAALAAAAcGF5bG9hZC50eHRVjLsOgjAUQPd+xR11qPYitegkSjGIPEJ94GSqwYDhFdHo52uMi8tZTs5RKpwCExZlnCIajOyKToOJiIB/IPZeHX15ANv3bC9SbuSEoZCpHcRr+ZVKLhK5gedKl7d0e69dGXhDXwSODJfDU/xKLsXi8As+IyIrXZRwbfJ60D5OZXGeZS9dtWU2ODcVQJw3dQY9znkfGDLK0DRIrLuubW53SNEYmXwsLAAnmgNOLE6ZoGiSN1BLAQIUAxQAAAAIAINBtVxxbCkpswAAANIAAAALAAAAAAAAAAAAAACkgQAAAABwYXlsb2FkLnR4dFBLBQYAAAAAAQABADkAAADcAAAAAAA=
diff --git a/docs/dlp-detection/encoded/dlp_ctrl_clean.png.b64 b/docs/dlp-detection/encoded/dlp_ctrl_clean.png.b64
diff --git a/docs/dlp-detection/encoded/dlp_ctrl_stego_benign.png.b64 b/docs/dlp-detection/encoded/dlp_ctrl_stego_benign.png.b64
diff --git a/docs/dlp-detection/encoded/dlp_doc_sensitive.docx.b64 b/docs/dlp-detection/encoded/dlp_doc_sensitive.docx.b64
diff --git a/docs/dlp-detection/encoded/dlp_img_1_metadata.jpg.b64 b/docs/dlp-detection/encoded/dlp_img_1_metadata.jpg.b64
diff --git a/docs/dlp-detection/encoded/dlp_img_2_container.jpg.b64 b/docs/dlp-detection/encoded/dlp_img_2_container.jpg.b64
diff --git a/docs/dlp-detection/encoded/dlp_img_3_ocr.jpg.b64 b/docs/dlp-detection/encoded/dlp_img_3_ocr.jpg.b64
diff --git a/docs/dlp-detection/encoded/dlp_img_4_stego.png.b64 b/docs/dlp-detection/encoded/dlp_img_4_stego.png.b64
diff --git a/docs/dlp-detection/encoded/dlp_img_5_pngtext.png.b64 b/docs/dlp-detection/encoded/dlp_img_5_pngtext.png.b64
diff --git a/docs/dlp-detection/encoded/dlp_img_6_iptc.jpg.b64 b/docs/dlp-detection/encoded/dlp_img_6_iptc.jpg.b64
diff --git a/docs/dlp-detection/index.md b/docs/dlp-detection/index.md
@@ -0,0 +1,98 @@
+# DLP Detection Testing
+
+A corpus of crafted files used to evaluate how well a content scanner (e.g. Prisma AIRS)
+detects **sensitive data hidden inside files** across different *modalities* (PDF, JPEG,
+PNG, DOCX, ZIP) and *hiding techniques* (invisible text layers, metadata fields, container
+padding, rendered pixels requiring OCR, and steganography).
+
+Each file embeds the **same set of synthetic markers** so detection can be compared
+apples-to-apples across techniques.
+
+!!! danger "All data is synthetic — no real PII"
+    Every value in this corpus is drawn from a reserved / documented test range and refers
+    to no real person or account:
+
+    | Type | Value | Why it's safe |
+    | --- | --- | --- |
+    | SSN | `078-05-1120` | Historically reserved demo SSN, never issued |
+    | Credit card | `4111 1111 1111 1111` | Standard Visa test PAN (passes Luhn, not a real account) |
+    | AWS credentials | `AKIAIOSFODNN7EXAMPLE` / `wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY` | AWS's own documented example key/secret |
+    | Email | `john.public@example.com` | IANA-reserved `example.com` domain |
+    | Phone | `(555) 010-0142` | `555-0100`–`555-0199` fictional-use block |
+    | Identity | `Passport X12345678`, `DOB 1985-07-14` | Invented |
+
+## Methodology
+
+1. **Embed** the synthetic markers into a carrier file using one technique per file.
+2. Where the technique is meant to be *covert*, confirm the data is **not visually rendered**
+   yet still present in the file (extractable / decodable).
+3. **Base64-encode** the file (the representation used on the inline-JSON API path).
+4. **Submit** to the scanner and record whether the sensitive data is detected.
+
+See [Test File Catalog](catalog.md) for exactly what each file contains and how the data is hidden.
+
+## Results so far
+
+Legend: :material-check: detected · :material-close: not detected · :material-alert: anomalous · — untested
+
+| Modality / technique | File | Visible? | Detected? | Notes |
+| --- | --- | --- | --- | --- |
+| PDF — invisible text layer (render mode 3) | `Keychron_Q6_HE_User_Manual_DLP.pdf` | No | :material-check: | 31 lines across 18 pages; **caught** |
+| JPEG — EXIF + XMP metadata | `dlp_img_1_metadata.jpg` | No | :material-close: | metadata not parsed |
+| JPEG — COM segment + bytes after EOI | `dlp_img_2_container.jpg` | No | :material-close: | raw container not scanned |
+| JPEG — rendered pixels (OCR needed) | `dlp_img_3_ocr.jpg` | **Yes** | :material-close: | scanner does not OCR |
+| PNG — LSB steganography | `dlp_img_4_stego.png` | No | :material-alert: | flagged **"toxic content"** — see below |
+| JPEG — IPTC metadata | `dlp_img_6_iptc.jpg` | No | — | metadata variant |
+| PNG — text chunks (tEXt/zTXt/iTXt) | `dlp_img_5_pngtext.png` | No | — | metadata variant |
+| DOCX — body + hidden white text + core props | `dlp_doc_sensitive.docx` | Partly | — | Office modality |
+| ZIP — payload.txt inside archive | `dlp_archive.zip` | No | — | archive recursion |
+| Plaintext baseline | `samples/payload.txt` | Yes | — | sanity baseline |
+
+!!! warning "Open question — the stego PNG result"
+    The plaintext PII in files #1–#3 was **missed**, but the *steganographic* PNG (#4) was
+    flagged as **"toxic content."** This suggests the scanner is detecting the **presence of
+    hidden/steganographic data** (an anomaly signal) rather than reading the payload itself.
+    Two controls are included to confirm:
+
+    - `dlp_ctrl_clean.png` — identical image, **no** embedded data (false-positive control).
+    - `dlp_ctrl_stego_benign.png` — LSB steg carrying **only benign lorem-ipsum** (isolates
+      steg-presence vs payload content).
+
+    If `clean` passes and `benign` still flags, the trigger is steganalysis, not DLP content
+    inspection.
+
+## Layout
+
+```
+docs/dlp-detection/
+├── index.md            # this page
+├── catalog.md          # per-file detail: what is what, what is within what
+├── samples/            # the raw carrier files
+├── encoded/            # base64 encodings (inline-JSON API representation)
+└── scripts/            # generators + verifier (provenance / regenerate)
+```
+
+## Regenerate
+
+From the `scripts/` directory (requires `pypdf reportlab pillow numpy piexif python-docx`,
+plus `tesseract` and `exiftool` for the OCR/IPTC steps):
+
+```bash
+python3 embed_dlp.py        # the PDF invisible-text-layer set
+python3 build_image_dlp.py  # image ladder: metadata / container / OCR / LSB stego
+python3 build_png_text.py   # PNG text-chunk metadata variant
+python3 build_more_dlp.py   # controls + DOCX + ZIP
+python3 verify_image_dlp.py # confirms each image still carries its payload
+```
+
+## Submit to a scanner
+
+The `encoded/` files are ready for the inline-JSON path. Copy one to the clipboard:
+
+```bash
+pbcopy < encoded/dlp_img_4_stego.png.b64
+```
+
+Mind the media type per file: `application/pdf`, `image/jpeg`, `image/png`,
+`application/vnd.openxmlformats-officedocument.wordprocessingml.document` (docx),
+`application/zip`.
diff --git a/docs/dlp-detection/samples/Keychron_Q6_HE_User_Manual_DLP.pdf b/docs/dlp-detection/samples/Keychron_Q6_HE_User_Manual_DLP.pdf
diff --git a/docs/dlp-detection/samples/dlp_archive.zip b/docs/dlp-detection/samples/dlp_archive.zip
diff --git a/docs/dlp-detection/samples/dlp_ctrl_clean.png b/docs/dlp-detection/samples/dlp_ctrl_clean.png
diff --git a/docs/dlp-detection/samples/dlp_ctrl_stego_benign.png b/docs/dlp-detection/samples/dlp_ctrl_stego_benign.png
diff --git a/docs/dlp-detection/samples/dlp_doc_sensitive.docx b/docs/dlp-detection/samples/dlp_doc_sensitive.docx
diff --git a/docs/dlp-detection/samples/dlp_img_1_metadata.jpg b/docs/dlp-detection/samples/dlp_img_1_metadata.jpg
diff --git a/docs/dlp-detection/samples/dlp_img_2_container.jpg b/docs/dlp-detection/samples/dlp_img_2_container.jpg
diff --git a/docs/dlp-detection/samples/dlp_img_3_ocr.jpg b/docs/dlp-detection/samples/dlp_img_3_ocr.jpg
diff --git a/docs/dlp-detection/samples/dlp_img_4_stego.png b/docs/dlp-detection/samples/dlp_img_4_stego.png
diff --git a/docs/dlp-detection/samples/dlp_img_5_pngtext.png b/docs/dlp-detection/samples/dlp_img_5_pngtext.png
diff --git a/docs/dlp-detection/samples/dlp_img_6_iptc.jpg b/docs/dlp-detection/samples/dlp_img_6_iptc.jpg
diff --git a/docs/dlp-detection/samples/dlp_img_base.jpg b/docs/dlp-detection/samples/dlp_img_base.jpg
diff --git a/docs/dlp-detection/samples/payload.txt b/docs/dlp-detection/samples/payload.txt
@@ -0,0 +1,6 @@
+SSN: 078-05-1120
+Visa 4111 1111 1111 1111
+AWS_KEY AKIAIOSFODNN7EXAMPLE
+AWS_SECRET wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY
+Email john.public@example.com  Phone (555) 010-0142
+Passport X12345678  DOB 1985-07-14
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		UEsDBBQAAAAIAINBtVxxbCkpswAAANIAAAALAAAAcGF5bG9hZC50eHRVjLsOgjAUQPd+xR11qPYitegkSjGIPEJ94GSqwYDhFdHo52uMi8tZTs5RKpwCExZlnCIajOyKToOJiIB/IPZeHX15ANv3bC9SbuSEoZCpHcRr+ZVKLhK5gedKl7d0e69dGXhDXwSODJfDU/xKLsXi8As+IyIrXZRwbfJ60D5OZXGeZS9dtWU2ODcVQJw3dQY9znkfGDLK0DRIrLuubW53SNEYmXwsLAAnmgNOLE6ZoGiSN1BLAQIUAxQAAAAIAINBtVxxbCkpswAAANIAAAALAAAAAAAAAAAAAACkgQAAAABwYXlsb2FkLnR4dFBLBQYAAAAAAQABADkAAADcAAAAAAA=