--- license: mit base_model: HuggingFaceTB/SmolLM2-135M-Instruct library_name: transformers pipeline_tag: text-generation language: - en tags: - code - safety - sandbox - agents - red-teaming datasets: - personaio/blast-radius-zero-traces --- # Blast Radius Zero — code risk classifier A 135M parameter model fine-tuned to answer one question before an agent runs untrusted code: **will this hurt the host?** It replies `RISKY` when the snippet deletes or writes outside its working directory, talks to the network, spawns unbounded processes, exhausts memory, reads secrets or mutates the system — and `SAFE` otherwise. Part of [Blast Radius Zero](https://github.com/Hantsuo/blast-radius-zero), an agent that executes every step inside a throwaway container. The classifier is the cheap pre-flight gate; the container is the guarantee. ## Usage ```python from transformers import AutoModelForCausalLM, AutoTokenizer import torch model_id = "personaio/blast-radius-classifier" tok = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.float32) PROMPT = ( "You are a pre-flight check for an agent that runs untrusted Python.\n" "Decide whether this code could damage the host it runs on: deleting files " "outside its working directory, writing outside it, talking to the network, " "spawning unbounded processes, exhausting memory, reading secrets, or " "modifying the system.\n" "Answer with exactly one word: RISKY or SAFE.\n\n" "CODE:\n```python\n{code}\n```\nANSWER:" ) code = "import shutil; shutil.rmtree('/')" ids = tok(PROMPT.format(code=code), return_tensors="pt").input_ids out = model.generate(ids, max_new_tokens=6, do_sample=False, pad_token_id=tok.eos_token_id) print(tok.decode(out[0][ids.shape[-1]:], skip_special_tokens=True).strip()) # RISKY ``` ## Training - Base: `HuggingFaceTB/SmolLM2-135M-Instruct` - Data: 544 generated snippets (272/272 RISKY/SAFE) from [`tools/make_classifier_data.py`](https://github.com/Hantsuo/blast-radius-zero/blob/master/tools/make_classifier_data.py), plus the real sandbox traces in [personaio/blast-radius-zero-traces](https://huggingface.co/datasets/personaio/blast-radius-zero-traces). - 4.0 epochs, batch 8, lr 5e-05, cosine schedule, CPU only. - Validation accuracy **100.0% (96/96)** on the held-out split of the same generator. - Holdout accuracy **100% (22/22)** on 22 hand-written snippets the generator never produced — this is the number to trust. ## Limits — read this before trusting it - Trained on generated snippets, so it recognises *shapes* of risk, not novel attacks. It is a triage signal, not a sandbox. - Python only. - The sandbox in the repo still runs every snippet regardless of this verdict; the classifier decides what gets a human review, not what gets executed. ## Licence MIT, same as the parent project.