Purpose

mg-exploitgen generates a CVE-driven exploit project scaffold. Given a CVE identifier, an operator-supplied CVE description, and a target-environment JSON, it lays down a Rust project skeleton ready to be instrumented and tested in a lab.

Inputs

  • <cve> — CVE identifier (case-insensitive; normalized to upper case; rejected if it contains anything outside [A-Za-z0-9-]).
  • --cve-description <path> — Markdown or text file with the CVE description. Not fetched from the network; the operator pastes it.
  • --target-env <path> — JSON describing the target stack, mitigations, reachability, and constraints.

Output

engagements/<name>/exploits/<CVE-ID>/
|-- Cargo.toml
|-- runbook.md
|-- src/
|   |-- main.rs         # thin orchestrator
|   |-- scanner.rs      # version detection
|   |-- validator.rs    # precondition check that doesn't fire payload
|   |-- payload.rs      # primitive stub + scope/legality comments
|   `-- cleanup.rs      # revert artifacts the exploit leaves
`-- tests/
    `-- smoke.rs        # scaffold compile-only test

Every generated runbook.md opens with:

Authorized testing only. Confirm scope before running.

The generated Cargo.toml uses a deterministic crate name derived from the CVE id: CVE-2026-1234 → exploit_cve_2026_1234.

CLI

mg-exploitgen scaffold acme-bounty CVE-2026-0001 \
    --cve-description /tmp/cve-2026-0001.md \
    --target-env /tmp/target-env.json

# Offline mode — deterministic placeholders, no LLM call
mg-exploitgen scaffold acme-bounty CVE-2026-0001 \
    --cve-description /tmp/cve-2026-0001.md \
    --target-env /tmp/target-env.json \
    --offline --force

How it works

  • Directory tree and Cargo skeleton come from static templates embedded via include_str!. The LLM never picks file names or controls directory structure.
  • The model only fills guidance fields that get pasted into file-header comments and into a numbered runbook step list:
    • scanner_strategy, validator_strategy, payload_strategy, cleanup_strategy
    • target_env_summary, cve_summary
    • runbook_steps, notes
  • The model returns one JSON object. The harness’s extract_first_json_object parser tolerates assistant chatter around the JSON block.
  • CVE description and target-env JSON are wrapped as untrusted <cve_description> / <target_env> evidence in the prompt.

Safety

  • No network access. CVE descriptions are operator-supplied paste-only.
  • Generated scaffold compiles cleanly under cargo check (verified with a smoke run during S4). The stage stubs deliberately do nothing until the operator implements them.
  • The “Authorized testing only” banner is hard-coded in the runbook template; you cannot generate a runbook without it.

Harness

mg-harness exposes exploit.scaffold (ReadOnly — no traffic leaves the box).