Skip to content

manta extract

Extract a payload from a model file. Supports safetensors and GGUF formats (auto-detected). Reverses the embed pipeline and fails closed if any later carrier tensor is missing or corrupted.

Usage

manta extract -m <MODEL> -o <OUTPUT> -t <TENSORS> [-d <DEPTH>] [-k <KEY>] [-e]

Options

Flag Short Default Description
--model -m required Model file containing payload (safetensors or GGUF)
--output -o required Output file for extracted payload
--tensors -t required Comma-separated tensor names (from extraction key)
--lsb-depth -d 3 LSB depth used during embedding
--key -k none Decryption passphrase
--ecc -e false Whether ECC was used during embedding

Pipeline

Extraction reverses the embed pipeline:

[LSB extract] → [MANT unframe] → [AES-256-GCM decrypt] → [ECC decode] → payload

Examples

Basic extraction

manta extract \
  -m weaponized.safetensors \
  -o payload.bin \
  -t "model.layers.0.mlp.down_proj.weight" \
  -d 3

Full pipeline (encrypted + ECC)

manta extract \
  -m weaponized.safetensors \
  -o payload.bin \
  -t "model.layers.0.mlp.down_proj.weight,model.layers.1.mlp.down_proj.weight" \
  -d 3 \
  -k "operator-passphrase" \
  -e

Errors

Error Cause
Frame magic not found Wrong tensor names, wrong depth, or no payload embedded in the first carrier tensor
Partial extraction detected An earlier carrier decoded successfully, but a later carrier tensor was missing or corrupted
Decryption failed Wrong passphrase
ECC decode failed Too many shards missing or zeroed (ECC recovers dropped tensors, not bit-level corruption — use --key for integrity)

Tip

The tensor names and order must match what was used during embedding. Use the extraction key JSON from the embed command.

See Also