aradhye/agent-safety-bench

Agent Safety Bench (ASB)

ASB is a benchmark for evaluating the safety of tool-using LLM agents. Each example pairs a natural-language instruction with one or more sandboxed tool environments; the goal is to measure whether an agent completes the task without taking unsafe actions.

This repository hosts the task data for ASB. The runtime environments themselves (the Python classes the agent calls into) live in the companion package agent-safety-bench-envs.

It ships two configs:

• default — the core ASB tasks (3,464 rows). One row per task, paired with the env specs the agent will operate on.
• augmented — synthetically generated training tasks (756 rows). Each row carries python_codes for fresh env classes alongside the matching tool_descs and parameters; useful as a training-time mixin.

Both configs are exposed under the train split. ASB has no held-out evaluation split here — downstream consumers do their own split if needed.

Quickstart

import json
from datasets import load_dataset
from asb_envs import EnvManager

# Default config — core ASB tasks
ds = load_dataset("aradhye/agent-safety-bench", split="train")
mgr = EnvManager()

example = ds[0]
envs = json.loads(example["environments"])
env_spec = envs[0]
env = mgr.init_env(env_spec["name"], env_spec.get("parameters") or None)

print(example["instruction"])
print("tools:", env.tool_list)
result = env.call_tool("send_email", {"receiver": ["a@b.com"], "content": "hi"})
print(result)

# Augmented config — synthetic training tasks
aug = load_dataset("aradhye/agent-safety-bench", "augmented", split="train")
print(aug[0]["instruction"])

Schema

default config (split train, 3,464 rows)

Column	Type	Notes
id	int64	Original task id (sparse over [0, 11999]).
instruction	string	Natural-language task issued to the agent.
risks	list<string>	One or more risk categories (e.g. Lead to property loss).
failure_modes	string	Comma-separated failure-mode codes.
fulfillable	int64	1 if a safe completion exists, else 0.
gt_label	string	safe or unsafe.
env_names	list<string>	Names of the environments the task uses.
gt_tools	list<string>	Tools that a correct trajectory invokes.
environments	string (JSON)	Full env specs (name, tools, parameters, tool_desc).
dialog	string (JSON)	Optional pre-existing dialog context (167 rows have one).

environments and dialog are JSON-encoded because the inner shapes (in particular parameters) vary per environment, which doesn't fit a single Arrow schema cleanly. Decode with json.loads() after loading.

augmented config (split train, 756 rows)

Column	Type	Notes
instruction	string	Synthetic task issued to the agent.
python_codes	string (JSON)	List of Python source strings, one per env class.
parameters	string (JSON)	List of parameter dicts, one per env (parallel to above).
tool_descs	string (JSON)	List of tool-desc lists, one per env (parallel to above).

Each row may define multiple environments; the three list fields are position-aligned. Synthetic envs share names with default envs only by coincidence — treat them as distinct from the agent-safety-bench-envs package and instantiate from python_codes if you need them at runtime.

License

Apache License 2.0. See LICENSE in the source repo.