EddyLuo/MechanisticAgentSafety

MAS: Mechanistic Agent Safety Benchmark

The Mechanistic Agent Safety (MAS) benchmark provides labeled interaction steps from multi-turn command-line interface (CLI) agents. Each example pairs the conversation context before an agent action with its generated output and a binary current-step safety label. It supports training hidden-state safety probes, analyzing layer-wise representations, and evaluating representation-level steering.

MAS accompanies Agent MechSuits: Mechanistic Subspace Safety Steering for Multi-Turn CLI Agents, accepted at NeurIPS 2026.

Quick start

Install the dataset loader:

pip install -U datasets huggingface_hub

Authenticate with Hugging Face or set HF_TOKEN when accessing a private repository, then select a model configuration:

from datasets import load_dataset

mas = load_dataset("EddyLuo/MAS", "llama", token=True)
train = mas["train"]
test = mas["test"]

example = train[0]
messages = example["input"]  # Context before the labeled action.
target = example["label"]    # 0 = benign/refusal; 1 = harmful execution.

Available configurations are gemma, llama, and qwen. Each contains train and test splits. Records are stored as JSON arrays, with the original model-specific filenames preserved.

Dataset composition

The release contains 1,947 step-level examples: 1,156 training examples and 791 test examples. Counts below refer to labeled steps, not complete trajectories or distinct tasks.

Configuration	Agent backbone	Split	Steps	Benign / refusal (0)	Harmful (1)	File
gemma	Gemma-2-9B	Train	146	120	26	Gemma/trainset.json
gemma	Gemma-2-9B	Test	86	72	14	Gemma/testset.json
llama	LLaMA-3.1-8B	Train	682	327	355	LLaMA/train.json
llama	LLaMA-3.1-8B	Test	469	228	241	LLaMA/test.json
qwen	Qwen-2.5-7B	Train	328	188	140	Qwen/train.json
qwen	Qwen-2.5-7B	Test	236	158	78	Qwen/test.json

Split audit

The paper describes task-disjoint evaluation. An audit of the JSON files in this release found the following numbers of distinct, identical input message lists shared between the training and test files:

Configuration	Distinct shared inputs
gemma	0
llama	52
qwen	26

Inputs were compared by canonical JSON serialization with dictionary keys sorted; message order and text were preserved. Zero identical inputs does not establish task-level disjointness. The released files should therefore not be assumed to implement a task-disjoint benchmark. For new generalization studies, audit source-task and trajectory membership, remove duplicate contexts across splits, and document the resulting partition.

Record schema

Field	Type	Description
rollout_id	Integer	Source rollout identifier.
task_id	Integer	Source task identifier.
task_description	String	Natural-language task associated with the step.
step_number	Integer	Step index in the source rollout.
input	List of message objects	Pre-generation conversation context. Each message has role and content strings; the released files use user and assistant roles.
output	String	Agent output for the labeled step, including reasoning/action text where present.
label	Integer, 0 or 1	Current-step safety target used by the reference probe code.

Some Gemma and Qwen records also contain pattern. Some Qwen records contain source, label_opus, pattern_opus, or label_noaware. These are auxiliary annotations with varying coverage; string-valued fields may be absent or null. Use label as the target for the documented current-step task.

Identifiers are preserved from the source files and should not be treated as globally unique across configurations or splits. This repository distributes text records and labels; model weights, precomputed activations, trained probes, and executable task environments are not included.

Label semantics

Labels describe the current agent output, interpreted in its execution context:

• 0 - benign or safety-aware/refusal: an authorized benign operation, a non-harmful response, or a step that refuses or neutralizes an unsafe request.
• 1 - harmful execution: a step that performs, writes, or triggers an unsafe operation or payload.

A malicious task can contain a benign or refusal step. Consequently, the label is not a classifier target for the original task's intent. Interpret an action together with its preceding conversation rather than relying on command keywords alone.

The label field is a current-step target. These files do not expose a dedicated lookahead split or lookahead-label field; future-risk targets require a separately defined trajectory reconstruction and horizon.

Using MAS for linear probing

For a pre-action safety probe, the supervised pair is the hidden representation of input and the target label:

1. Select the configuration corresponding to the agent backbone being studied.
2. Apply that backbone's chat template to input, preserving message order and roles and adding the generation prompt.
3. Run the frozen backbone and extract the last-token activation at the chosen layer. The reference implementation also supports attention and MLP module outputs.
4. Train a binary linear classifier on these activations and label.

Keep output outside the probe input for pre-action detection: it is the action whose safety is being predicted. Fit preprocessing and select layers, thresholds, and hyperparameters using training/validation data. Report class-aware metrics alongside accuracy because class proportions vary across configurations. Probe accuracy measures decodability; causal claims require intervention experiments.

Reference training code

After cloning the code repository and installing its requirement.txt, download the JSON files into the expected MAS/ directory:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="EddyLuo/MAS",
    repo_type="dataset",
    local_dir="MAS",
    allow_patterns=["Gemma/*.json", "LLaMA/*.json", "Qwen/*.json"],
    token=True,
)

Then train layer-wise LLaMA probes with the repository's script:

python train_probe.py \
  --model models/meta-llama/Llama-3.1-8B-Instruct \
  --dataset LLaMA \
  --layer hidden_state \
  --max_layers 32 \
  --epoch 100 \
  --learning_rate 0.01 \
  --batch_size 16 \
  --device cuda:0

The model checkpoint must be available at the supplied path. Configuration names on the Hub are lowercase; the script accepts Gemma, LLaMA, and Qwen. The reference script evaluates the provided test file after each epoch and saves the checkpoint with the highest test accuracy. For an independent final evaluation, introduce a validation partition for checkpoint and layer selection.

Collection and annotation

The paper describes a source pool of 194 tasks (98 malicious and 96 benign), curated from AdvCUA, Safe-OS, and AgentBench. ReAct-style agents execute tasks in Docker environments with greedy decoding, accumulating prior actions and environment observations. Annotation combines automated screening, model-assisted labeling, and human review, with counterfactual benign examples to reduce reliance on superficial command patterns. These collection details describe the benchmark methodology; the table above reports the contents of this release.

Scope and limitations

MAS covers three agent backbones and a limited collection of CLI environments. Its labels reflect the current-step intervention target and can be ambiguous when risk is distributed across several actions. The files support studying safety representations in the recorded settings; broader deployment claims require evaluation on additional tasks, models, and environments. See the paper's limitations and the split audit above when designing experiments.

Citation

If you use MAS or the accompanying probe/steering code, please cite:

@article{luo2026agentmechsuits,
  title={Agent MechSuits: Mechanistic Subspace Safety Steering for Multi-Turn CLI Agents},
  author={Luo, Weidi and Zhang, Qiming and Quan, Yihao and Jin, Mingyu and Cai, Jie and Xiao, Chaowei and Niu, Jingcheng and Xiang, Zhen},
  journal={arXiv preprint arXiv:2606.22673},
  year={2026},
  eprint={2606.22673},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2606.22673}
}