javyduck/agentweaver

ArenaWeaver Prompt-Injection Benchmark

A collection of legitimate agent tasks, saved prompt injections, and executable application data for studying whether agents respect user intent when working with untrusted content.

The dataset accompanies ArenaWeaver: Automatic Generation of Environments and Attacks for Prompt Injection Evaluation at Scale. Cases span browser interaction, Model Context Protocol (MCP) tools, and workflows combining both across 40 reconstructed applications. The companion AgentWeaver repository provides the environments and evaluation runtime; access to the code repository is separate from access to this public dataset.

Choose a collection

Configuration	Cases	Intended use
eval — default	660	The curated benchmark used for the paper's model comparisons. Start here to evaluate a model.
all	2,232	All cases recorded as successful during the original attack search, including the 660 benchmark cases.

Each configuration has one test split. eval is a subset of all, not a separate train/test partition. Original search success does not imply that an attack will succeed against a new model. Because unsuccessful searches are excluded, this collection cannot estimate the success rate of the original search campaign.

Scenario	eval	all
One application, MCP tools	191	1,001
One application, browser UI	120	544
Multiple applications, MCP tools	129	292
Multiple applications, browser UIs	100	156
Mixed browser and MCP access	120	239

Browse a case

The Dataset Viewer offers one row per case, with the task and attack shown first:

Column	Meaning
benign_task	The legitimate instruction given to the evaluated agent.
attacker_goal	The unauthorized outcome the attacker tries to induce.
applications	Applications and interfaces involved in the case.
scenario_name	One of the five workflow types listed above.
injection	The saved sequence of attack actions, including their payload arguments.

The remaining columns preserve the original case metadata and executable-file references. Select eval or all to switch collections. The same rows can be loaded in Python:

from datasets import load_dataset

cases = load_dataset("javyduck/agentweaver", "eval", split="test")

These Parquet rows under preview/ are a browsing aid, not a replacement for the executable inputs. Running a case also requires its saved databases and any referenced attachments.

Clean and attacked runs use the same user task

	Clean run	Attacked run
Instruction given to the evaluated agent	The legitimate user task	The same legitimate user task
Application state	Clean initial data	Data containing the saved injection
Purpose	Measure legitimate task completion	Measure whether the injection causes the attacker goal to be achieved

The attacker goal is evaluation metadata, not a user instruction given to the evaluated agent. In an attacked run, the agent may encounter the injected content while following its own action sequence. Shared clean databases can support several tasks; the corresponding attacked snapshots retain each case's intervention.

What the download contains

The original case indexes, data/eval.jsonl and data/all.jsonl, map cases to their executable files. artifacts/native-*.tar.gz stores those files in shared archives; data/native-files.jsonl and data/shards.jsonl record locations and integrity hashes.

File	Contents
attack_config.yaml	Legitimate task, attacker goal, and case configuration.
persona.yaml	Scenario persona and application/interface assignments.
data/<application>/seed.db	Clean initial SQLite database for each application.
winning_db/<application>.db	Saved application database after inserting the attack.
winning_injection_chain.yaml	Ordered attack actions and their payload arguments.
Referenced attachments	Binary content used by an intervention, when applicable.

These paths occur within each case's native projects/ layout. The archives also retain the original search metadata needed by the evaluation runtime. They do not distribute victim trajectories, per-model judgments, or internal agent logs. An injection action can place several content items, so action counts and content-item counts are reported separately.

Download and evaluate

No Hugging Face login or token is needed to download this public dataset. Running models or a model-based judge requires the relevant provider credentials.

After installing the companion code following its setup instructions, run this from the code repository root:

python scripts/evaluate_benchmark.py --config eval

This downloads the pinned dataset release, verifies file hashes, and restores the selected inputs into a separate run directory. It does not start Docker or call a model. Use --config all for the full collection. Add --victims MODEL_ID --execute to run an evaluation after configuring model access; see the evaluation guide for supported model IDs, judge selection, and runtime requirements.

To download the files without the companion code:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="javyduck/agentweaver",
    repo_type="dataset",
    local_dir="agentweaver-data",
)

The included unpack.py can restore the native layout into a new directory. It checks archive and file hashes and refuses to overwrite differing existing files:

mkdir -p benchmark-inputs
python agentweaver-data/unpack.py --config eval --destination ./benchmark-inputs

Scope and limitations

The applications and account data are reconstructed test environments, not live service accounts. Results can vary with model versions and sampling; the collection is not a guarantee of reproducing a historical model score.

All 660 eval cases include historical attacked snapshots. One all-only Google Sheets browser case lacks its historical attacked database; its clean database and saved attack chain are included, but reapplying the chain cannot recover the original database bytes or generated IDs. Separately, four released cases have a task identity that conflicts with the saved state, including one eval case. The evaluation runtime rejects these identity mismatches rather than silently evaluating the wrong account. These limitations apply to the original downloadable inputs as well as the viewer.

License and safe use

Dataset: CC BY 4.0. Companion code: MIT.

The cases intentionally contain misleading or malicious instructions. Treat them as untrusted data and run them only in isolated test environments, never against live user accounts or services.