djain95/adaptive-adversaries

Adaptive Adversaries — Security Arena data

Version 1.0.0, September 11, 2026, accompanies Adaptive Adversaries: A Multi-Turn, Multi-LLM Benchmark for LLM Agent Security, by Devina Jain, David Hartmann and Chuan Li (arXiv:2607.18063).

Download dataset and reproduction bundle · Code · Versioned release and checksums

The benchmark measures scenario-defined security failures during adaptive cross-session evaluation. Attackers retain previous attempts and responses; the primary defender condition starts a fresh session per attempt. The competition adds participant-designed harnesses around a common backbone. Benign-task evaluation measures ordinary task performance separately.

Included evidence

Study	Records	Location
Frontier matrix, 21 scenarios × 3 attackers × 3 defenders × 5 battles	945 battles; 13,189 saved responses	transcripts/
Three-attacker targeted replication	540 battles	3x3_n=20/
Provenance paragraph A/B	540 battles (270 per arm)	studies/provenance/
Attacker-history control	72 battles	studies/history/
Defender-state 2×2 factorial	216 battles (54 per arm)	studies/defender_state/
July fixed-Opus precision study	540 memberships: 135 shared + 405 additional battles	studies/precision_n60/, studies/precision_added/
Frozen replay	62 source sequences; 558 trials	studies/replay/ and derived replay inputs
Frontier benign provenance pilot	60 responses	studies/frontier_benign/
Competition held-out evaluation	18,422 battles; two snapshots	competition_traces/battles.jsonl.gz
Competition development iterations	93,291 battles; 52 populated snapshots	competition_traces/public_iteration_battles.jsonl.gz
Competition benign evaluation	3,083 eligible events; 9,171 topics; 10,151 attempts	studies/competition_utility/
Reference tier	1,050 saved run summaries	baseline_vs_baseline_n=50/
Calibration, descriptive embedding and compute estimate inputs	Study-specific	calibration/, studies/calibration/, studies/embedding/, studies/compute/

The 135 shared IDs are listed explicitly in the precision membership manifest. The table's rows must not be summed as independent runs. The older 48-source replay and strict-scoring summary directories are historical evidence, identified separately by the release manifest; they do not supply the revised replay or primary metric. Cached bytecode, confirmed duplicate summaries and _archive/ are excluded from the default export and remain flagged locally.

Reproduce the paper

Install the matching code package with its analysis dependencies, then run:

adaptive-adversaries-reproduce --data /path/to/this/dataset --out reproduced

No model API calls are made. The default primary profile, policy_compliance_v1, gives 108/945 frontier successes: Opus 25/315, GPT-5.4 30/315, Gemini 53/315. Frozen strict checks give 78; broader exact-output checks give 152; the corresponding repaired-code sensitivities give 97 and 141. These are separately named criteria. Archived winner labels give 121 and are not overwritten.

Defender-state arms A/B/C/D give 31/17/2/8 out of 54. At fixed attacker instructions, defender memory reduces ASR by 25.9 percentage points (Fisher p=0.0114566). D changes both state and attacker instructions relative to A. Competition primary security uses 3,439/17,584 available battles after excluding 838 infrastructure failures. Competition benign performance is 8,519/9,171 at the first attempt and 8,934/9,171 after retries; team medians are 96.6% and 99.1%.

Prepared derived outputs are under derived/1.0.0/, unchanged from the validated candidate. They include profile-specific outcomes and first-success rounds. release_manifest.json records distributed files, hashes and roles. croissant.json describes the record sets; SCHEMA.md defines their fields.

Load local records

from datasets import load_dataset

frontier = load_dataset(
    "json", data_files="/path/to/dataset/transcripts/*.jsonl", split="train"
)

A dependency-free loader is also sufficient:

import gzip, json
with gzip.open("/path/to/dataset/studies/defender_state/records.jsonl.gz", "rt") as f:
    battles = [json.loads(line) for line in f]

from adaptive_adversaries.scoring import score
won, reason = score("tickettriage", response_text, "policy_compliance_v1")

The bare scenario-module import previously shown in this card was incorrect; use the installed, versioned scoring API. Historical source copies in scenarios/ are not the primary scoring API.

Scope, provenance and use

Competition IDs are Baseline and T01–T21; there are 22 attacker and 21 defender identities, with T15 lacking an active held-out defender. Registration tokens and source IDs in utility data are opaque digests. Neither the private identity map nor participant harness source is distributed. Utility/security joins describe historical teams, not proven identical harness revisions.

Saved attack text is untrusted benchmark data. Evaluate it inside the benchmark's defined scenario interface. Scenario successes are operational checks, not a guarantee of real-world compromise or complete semantic correctness. Zero observed successes is not a proof of robustness.

Missing provenance is explicit: replay target responses; frontier-pilot questions, complete rendered inputs and judge explanations; exact embedding text IDs/encoder revision; and historical per-call usage, billing and complete runtime model revisions. The character-based compute reconstruction is not measured token usage.

The dataset is licensed under CC BY 4.0; code has a separate MIT license. The competition organizer confirmed that the participation terms or existing permissions cover publication of the included interaction traces and benign-task records. DATASHEET.md and RELEASE_BASIS.md document the release. Report issues through the project repository.

@misc{jain2026adaptiveadversaries,
  title={Adaptive Adversaries: A Multi-Turn, Multi-LLM Benchmark for LLM Agent Security},
  author={Devina Jain and David Hartmann and Chuan Li},
  year={2026},
  eprint={2607.18063},
  archivePrefix={arXiv},
  primaryClass={cs.CR}
}