yzhao062/agent-risk-logs

Agent Risk Experiment Logs

Overview

This dataset contains the public experiment logs, evaluation trajectories, and analysis records supporting the paper "agent-risk: Risk Assessment for Changes to AI Agents" (arXiv preprint; code repository: https://github.com/yzhao062/agent-risk). The logs document 2,495 analysis-sufficient files (3,305.33 MiB / 3,465.89 MB total uncompressed payload) capturing counterfactual replays, live stopping and retry gates, targeted continuation executions from saved container states, multi-turn bash sandboxes, model-update reruns on high-performance accelerators, and paired announcement tests across Qwen-series agent models.

Run Inventory and Paper Mapping

The table below maps the run codes used in file names, directory layouts, and execution manifests to the descriptive names used in the paper (aligned with the master display map in ARTIFACT.md).

Run Code	Name in the Paper	Directory Path	File Count	Size (MiB)
B0	The exploratory log	logs/pilot-ab/b0/	3	13.24 MiB
Arm A	The exploratory live gates	logs/pilot-ab/val2_t1/ to val2_t3/	9	27.16 MiB
B1	The held-out log	logs/arms-cd/b1/	4	24.28 MiB
Arm C	The live retry gate	logs/arms-cd/c_retry/	3	12.95 MiB
Arm D	The first announcement test	logs/arms-cd/d_announce/	3	9.12 MiB
Bounds	Held-out and pooled bounds, and the held-out log's serving record	logs/arms-cd/	6	0.20 MiB
D58	The registered replication (the replication log and the live replication runs)	logs/d58/	81	86.03 MiB
D64	The paired announcement test	logs/d64/	50	49.36 MiB
N2 stage 1	The step-0 continuations	logs/n2/stage1/, logs/n2/	9	3.10 MiB
D60	Held-out monitor scores recomputed (Table D.3; 'monitor rescoring' in logs)	logs/n2/d60/	1	1.05 MiB
Arm E	The model-update reruns	arm-e/	110	21.00 MiB
D65	The long-horizon log	logs/d65/	29	59.44 MiB
D66	The two-agent relay	logs/d66/	521	670.79 MiB
E1	The targeted-execution run (retry gate)	logs/e1/	342	275.71 MiB
E2	The stronger-agent run	logs/e2/	115	241.16 MiB
E3	The airline run	logs/e3/	102	118.99 MiB
E5	The airline retry gate by targeted execution	logs/e5/	48	42.84 MiB
E7	The two-gate run (two retry gates by targeted execution from one log)	logs/e7/	1059	1648.90 MiB
Total			2495	3305.33 MiB

File Formats and Inspect Archive Reading

The release includes JSON Lines logs (episodes.jsonl), pre-registered execution manifests (plan.json, run.json, started.jsonl), analysis records (analysis_*.json, prediction_*.json), and Inspect evaluation archives (.eval).

Inspect .eval files are standard zip archives whose members are compressed with Zstandard (zstd, compression method ID 93). Python's built-in zipfile module reads compression method 93 natively from Python 3.14 on; on earlier versions, register a decompressor from the zstandard package.

The following Python snippet registers zstandard with zipfile and reads member streams directly:

import zipfile, zstandard as zstd

_orig_check, _orig_decomp = zipfile._check_compression, zipfile._get_decompressor
zipfile._check_compression = lambda c: None if c == 93 else _orig_check(c)
zipfile._get_decompressor = lambda c: zstd.ZstdDecompressor().decompressobj() if c == 93 else _orig_decomp(c)

with zipfile.ZipFile("path/to/archive.eval") as z:
    for name in z.namelist():
        data = z.read(name)
        print(f"{name}: {len(data)} bytes")

Download Instructions

Download Using Python huggingface_hub

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="yzhao062/agent-risk-logs",
    repo_type="dataset",
    local_dir="data",
    allow_patterns=["logs/*", "arm-e/*", "MANIFEST.tsv", "REDACTIONS.tsv"],
)

Run this from the root of a clone of the GitHub repository. The patterns skip this dataset card, so the repository's own data/README.md stays unchanged; MANIFEST.tsv and REDACTIONS.tsv are byte-identical to the copies in the repository.

Download Using the Hugging Face CLI

# Using modern hf CLI:
hf download yzhao062/agent-risk-logs --repo-type dataset --local-dir data --include "logs/*" "arm-e/*" MANIFEST.tsv REDACTIONS.tsv

# Or using huggingface-cli:
huggingface-cli download yzhao062/agent-risk-logs --repo-type dataset --local-dir data --include "logs/*" "arm-e/*" MANIFEST.tsv REDACTIONS.tsv

Pointing Analysis Scripts to Downloaded Logs

To run the paper's analysis commands against the downloaded logs:

export AGENT_RISK_LOGS=$(pwd)/data/logs

Integrity Verification

The dataset includes a checksum manifest (MANIFEST.tsv) tabulating the relative path, exact byte size, and SHA-256 digest of every released file (excluding the dataset root metadata files README.md, MANIFEST.tsv, and REDACTIONS.tsv).

Verify download integrity with the following Python snippet:

import hashlib
from pathlib import Path

# When downloaded into data/ (or run within the dataset root):
data_root = Path("data") if Path("data").exists() else Path(".")
manifest_path = data_root / "MANIFEST.tsv" if (data_root / "MANIFEST.tsv").exists() else Path("MANIFEST.tsv")

mismatches, missing, total = [], [], 0
with open(manifest_path, "r", encoding="utf-8") as f:
    next(f)  # skip header
    for line in f:
        total += 1
        path_str, byte_str, expected_sha, *_ = line.strip().split("\t")
        target = data_root / path_str
        if not target.exists():
            missing.append(path_str)
            continue
        data = target.read_bytes()
        if len(data) != int(byte_str) or hashlib.sha256(data).hexdigest() != expected_sha:
            mismatches.append(path_str)

print(f"Verified {total} files: {total - len(missing) - len(mismatches)} OK, {len(missing)} missing, {len(mismatches)} mismatched.")

Redaction Policy

All released files have undergone systematic redaction to remove ephemeral cloud infrastructure identifiers, internal machine hostnames, account allocation numbers, and private local user paths, while preserving scientific data, agent decision sequences, and benchmark evaluation outputs:

• Public cloud VM IPv4 addresses are replaced with <vm-ip>.
• Internal cloud instance hostnames are replaced with vm-<n>.
• High-performance computing cluster node hostnames are replaced with <hpc-node>.
• Local user home paths are replaced with <home>, and bare usernames with <user>.
• Paths under /home/ubuntu/, the default user of the cloud VM image, are kept; they name no person or machine.
• Cloud project and allocation codes are replaced with <allocation>.
• SSH connection strings are replaced with ubuntu@<vm-ip>.
• Cloud instance UUIDs are replaced with <instance-id>.

Hash-Preserved Pre-Registered Files

Six pre-computed analysis and prediction records contain only a local home directory string in metadata or path records. These files remain byte-identical to their pre-registered versions so that independent evaluators can verify them against registered SHA-256 digests:

1. logs/arms-cd/retry_bounds_b0.json (SHA-256: f7ac1044a281d95daa7f1ff2cde5b3ada87545b619f520fb9fc3d41c6ccca789)
2. logs/arms-cd/retry_bounds_b1.json (SHA-256: 6410cca33585820917f84fc3b4122059dfca611cae98fe7dd0de4384a3a34a2f)
3. logs/arms-cd/retry_bounds_pooled.json (SHA-256: 73da02390af153bce4a23de04cb4b46d90fd931bfc692fd8e99a088c4c1c2102)
4. logs/e1/analysis_e1.json (SHA-256: 0c6dc48ff8cdd97d4fad553aa0c5da892e2569e099d6f7e330aa8a11942ce43a)
5. logs/e5/prediction_e5.json (SHA-256: 0413e2493c99313dc9eb22d09c75b048cad4530cd559e2c5989300ce7b4c8b54)
6. logs/e7/analysis_e7.json (SHA-256: 8c41d9a5335fd08de367d05adff11e55ef9d7672ad958cfb9f7ce4cba19ec09e)

Redacted Pre-Registration Launch Records

The launch copies of pre-registrations for runs E3, E5, and E7 were redacted using the standard redaction rules above to remove public VM floating IPs, hostnames, and instance codes. Their original SHA-256 digests remain on record in their respective launch records and pre-registrations, and the unredacted originals are available to reviewers upon request:

1. logs/e3/E3-PREREG-2026-10-05.launched.md:
   • Original committed SHA-256: d64860c863268eda669dfb2ddd0a8cfcb430643fc2d52e00d1b933d7bc9220d1 (committed in logs/e3/LAUNCH-REGISTRATION.txt)
   • Released redacted SHA-256: b582fead6df6de168ef437e26f402a0a2dd48294aea29d0fe7f944e66ba219a0
2. logs/e5/E5-PREREG-2026-10-06.launched.md:
   • Original committed SHA-256: 10e7923f20cae46f06e43b36773c5a8853e27eea06d721718f1c59c7d8d816e9 (committed in registrations/E5-PREREG-2026-10-06.md)
   • Released redacted SHA-256: dd9af5f352e69602c309bf8e5160b15cc2541ef65a8d4b1dcbd859c71baf166b
3. logs/e7/E7-PREREG-2026-10-06.launched.md:
   • Original committed SHA-256: 3302b3e92b9457b5fb8020b0ef4dd0e7a2fad3a46e85616941b44f6740d1f76f (committed in registrations/E7-PREREG-2026-10-06.md)
   • Released redacted SHA-256: 79bade77d8fde5e52964fb715950c79b2ec23f7ee728e29d0a6025f33cd6e6fb

Detailed redaction records for all 58 redacted files are published in REDACTIONS.tsv.

Excluded Runs

The following run codes represent development calibrations or runs outside this paper that are not included in the release:

• e4
• e6
• e8
• val1
• calib

Runs outside this paper are not included. (Note that bi_baseline and c1_baseline are inputs to Arm E, released under arm-e/data/.)

Licenses

• Experiment Logs Payload: The logs, trajectory traces, analysis outputs, and manifests in this repository are released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).
• Evaluation Tasks: Task text and bash environment specifications derive from ControlArena (UKGovernmentBEIS/control-arena), used under the MIT License. Airline dialog tasks in runs E3 and E5 derive from tau2-bench (sierra-research/tau2-bench), used under the MIT License.
• Model Generations: Agent and monitor outputs were generated using models in the Qwen family (Qwen3-4B-Instruct-2507, Qwen3-14B and its FP8 build, Qwen2.5-14B-Instruct, and Qwen3.5-27B-FP8), released under the Apache-2.0 License.

Citation

@article{zhao2026agentrisk,
  title   = {agent-risk: Risk Assessment for Changes to AI Agents},
  author  = {Zhao, Yue and Xu, Zixiang and Zhang, Angela Huyue and Yang, S. Alex and Hu, Xiyang},
  journal = {arXiv preprint},
  year    = {2026}
}