Dense optical flow
Compute a motion vector for every pixel between consecutive frames.
Ghost
Font
Decoder
Ghost fonts are marketed as text AI cannot read. This plugin pulls the hidden message out.
Output: decoded message
What it is
What is a ghost font?
A ghost font writes a message into motion, not into a single frame. In the real video beside this text, dots inside the letters drift against a background of dots moving another way.
Pause the video and the letters disappear. No single frame contains a visible difference, but an agent can measure the motion, rebuild the letter shapes, and return the exact hidden message.
Ghost-font video · original
Agent result · decoded in Codex
Process
How the decoder breaks it
The claim has one flaw: the message is the motion. The decoder reads that motion in four stages—no model weights and no GPU.
Subtract the background
Use median flow as the background, then isolate pixels moving against it.
Register the drift
Align every score map so drifting glyphs stack sharply instead of smearing.
Accumulate & read
Sum the signal, cancel the noise, clean the letters, and pass them to OCR.
The security claim, tested
They claim it's AI-proof. This plugin breaks through it easily.
Hiding a message in motion is a classic computer-vision problem, not a security boundary. Dense optical flow and phase correlation are decades-old tools.
A protection that collapses to a few hundred lines of Python is a visual novelty—not a defense.
Obfuscation isn't encryption.
Vision isn't a weakness—it's the model.
Vision isn't a weakness—it's the model.
01
Evidence 01
The claim — “AI can't read this”
Agents read video frame by frame, and every paused frame appears to be pure noise.
02
Evidence 02
The reality — AI reads motion too
The message lives in the relationship between frames, so the decoder reads that relationship.
03
Evidence 03
The verdict — A visual effect
Classic computer vision reveals the text reliably. Nothing cryptographic stays hidden.
Chat proof
The same message. Two ways in.
Four unedited chat captures show Codex and Claude recovering I LOVE YOU from the same ghost-font video.
Codex · reusable prompt
Decoded: I LOVE YOU
The decoder prompt is pasted directly into a regular Codex chat.
Claude · reusable prompt
Decoded: I LOVE YOU
The same standalone prompt runs in Claude and returns the matching result.
Codex · installed plugin
Decoded: I LOVE YOU
The installed Ghost Font Decoder plugin handles the full decode workflow.
Claude · installed skill
Decoded: I LOVE YOU
The installed /ghost-decode skill returns the same message in Claude.
Install
Give your agent the skill.
One repository, two agent plugins, a plain Python CLI, and a reusable chat prompt.
Claude Code
Install from the plugin marketplace.
/plugin marketplace add haroontrailblazer/ghost-font-decoder /plugin install ghost-font-decoder@ghost-font-tools
Codex
Install the plugin, then start a new task.
Use $ghost-decode to tell me what ghost-video.mp4 says.
Plain Python
Run the decoder without an agent.
pip install -r requirements.txt python decode.py examples/ghost-message.mp4
Any chat — no install
Paste the decoder prompt and attach a video. ChatGPT is not supported because its shorter execution timeout interrupts the decoder.
Open prompts/decode-in-chat.md Paste it into Claude or Codex.