hunter-crypto-7b
A local cryptography-attack specialist for authorized security work, by Naderu — a BytesBrains Pte. Ltd. venture.
hunter-crypto is a compact, specialised model for the cryptography corner of penetration
testing and CTF work: given a challenge — a cipher, a weak-parameter RSA key, an oracle — it
identifies the weakness and produces a runnable attack script (PyCryptodome / SageMath) that
recovers the plaintext or flag. It runs fully offline on Apple Silicon, so engagement data never
leaves the box. It is the first model of the Hunter family.
- Fine-tuned from:
Qwen/Qwen2.5-Coder-7B-Instruct (Apache-2.0)
- Method: QLoRA (r=8, scale=20) on a 4-bit MLX quant, assistant-tokens-only, 1000 iters, MLX on Apple Silicon
- This artifact: merged bf16 weights (adapter fused into the base, de-quantized)
- Full provenance, recipe, and eval harness: github.com/nandal/naderu —
models/hunter-crypto/
Intended use
A local specialist behind an authorized pentest / CTF workflow, handling the cryptography
corner offline: recovering plaintext/flags from cryptographically weak or misconfigured
constructions during authorized testing, CTF competition, and security education.
Authorized use only. This model is for authorized penetration testing, defensive security, and
CTF/education. It is not for attacking systems you do not have explicit permission to test. It
targets weak/misconfigured crypto for assessment; it is not a tool for defeating properly-deployed
modern cryptography, and does not claim to.
Out of scope: general chat, factual Q&A, non-crypto exploitation, and any use requiring
guaranteed correctness. Its output is a candidate attack to be run and verified, not an authority.
The task contract
Given a challenge (description + material: ciphertext, public parameters, an oracle interface,
source), the model returns (1) a short identification of the weakness, and (2) a single runnable
script that, when executed against the provided material, prints the recovered plaintext/flag.
Evaluation
Graded by end-to-end execution — the model's script must run and print the known secret (no partial
credit) — over a frozen held-out set verified disjoint from training. Base = the 4-bit Qwen2.5-Coder
base with no adapter; trained = this model.
| Tier | Metric | Gate | Base | Trained |
|---|
| In-distribution (48, six trained categories) | solve_rate | ≥0.95 → PASS | 0.333 | 1.000 |
| In-distribution (48) | valid_script_rate | ≥0.95 → PASS | 0.333 | 1.000 |
| Independent-method (36, different generator) | solve_rate | reported | 0.250 | 0.972 |
| OOD-hard (24, untrained RSA attacks) | solve_rate | non-gating | 0.000 | 0.125 |
Trained categories (all maxed on the in-distribution tier):
xor1,
xorK,
caesar,
affine,
base_chain,
rsa_low_e. The independent-method result (0.972 on a set built by a
different
generator with varied encodings/labels/parameters) is the signal that the model learned the attacks,
not just the training generator's surface templates. Full breakdown:
models/hunter-crypto/benchmarks.md.
Limitations, risks & safety
- Small, narrow model. It proposes attacks for weak crypto; it is not a cryptanalysis engine
and will be wrong on hard or out-of-distribution problems (see OOD tier). Always run and verify.
- No guarantee of correctness or safety of generated code — run generated scripts in a sandbox.
- Dual-use. Crypto-attack capability is inherently dual-use; scope is authorized assessment of
weak/misconfigured crypto, shipped as one corner of a human-consent-gated workflow.
License
Apache-2.0, inheriting Qwen/Qwen2.5-Coder-7B-Instruct's upstream terms.
