Ground-truth-labeled Linux server log and eBPF kernel event streams for cybersecurity continual learning research. Designed for use with the security-gym Gymnasium replay environment, which replays these streams without episode boundaries under agent-controlled defenses (blocklist, throttle list, isolation mode) that suppress future observations. The environment is trace-driven: the telemetry here is recorded, not generated by a dynamics model, so the loop is closed on observability but not on adversary behavior. Blocking an attacker suppresses that attacker's events; the recorded attacker does not change tactics in response.

This HuggingFace mirror hosts three of the four composed experiment streams (the 7-day smoke test, the 30-day mixed-attack stream, and the 90-day eBPF ablation stream, totaling ~2.8 GB compressed). The full release including the 365-day long-horizon stability stream (8.1 GB compressed) and the underlying base databases lives on Zenodo:

Concept DOI (always resolves to latest): 10.5281/zenodo.18901541
v4.1 version DOI (frozen for reproducibility): 10.5281/zenodo.21763493

Companion paper

Lawson, K. and Malik, H. (2026). Security-Gym: Evaluating Temporally-Uniform Agents on High-Fidelity Linux Telemetry. Continual Reinforcement Learning Workshop, Reinforcement Learning Conference (RLC) 2026, Montreal, August 15, 2026. PDF

The workshop is non-archival and publishes no proceedings, so the paper has no DOI and should not be cited as a conference publication. Cite the version-specific dataset DOI (10.5281/zenodo.21763493) when reporting results on Security-Gym streams.

Files in this mirror

File	Compressed	Decompressed	Events	Malicious	Campaigns	Duration
`exp_7d_brute_v4.db.zst`	156 MB	1.9 GB	4,891,541	25,980	SSH brute force only	7 days
`exp_30d_heavy_v4.db.zst`	686 MB	8.5 GB	21,511,208	609,520	Mixed, heavy attack rate	30 days
`exp_90d_v4.db.zst`	2.0 GB	25 GB	63,212,997	549,787	Mixed, moderate rate	90 days

Files only on Zenodo (not mirrored here due to size)

File	Compressed	Decompressed	Events	Malicious	Campaigns	Duration
`exp_365d_realistic_v4.db.zst`	8.1 GB	101 GB	257,654,256	1,857,178	Mixed, realistic rate	365 days
`benign_v4.db.zst`	434 MB	5.8 GB	11,159,241	0	(base benign)	—
`campaigns_v2.db.zst`	4.2 MB	41 MB	60,468	30,436	(base campaigns)	—

The 365-day stream is the long-horizon stability target used in the companion paper. Fetch it from Zenodo when running long-horizon evaluations; the three streams hosted here cover the 30-day mixed-attack stream used for the main prediction results, the 90-day eBPF ablation stream, and the 7-day smoke test.

Quick start

pip install security-gym huggingface_hub

from huggingface_hub import hf_hub_download
import zstandard, sqlite3

path = hf_hub_download(
    repo_id="j-klawson/security-gym-v4",
    filename="exp_30d_heavy_v4.db.zst",
    repo_type="dataset",
)

# Decompress to a local file
with open(path, "rb") as f, open("exp_30d_heavy_v4.db", "wb") as out:
    zstandard.ZstdDecompressor().copy_stream(f, out)

# Stream events through the Gymnasium environment
import gymnasium as gym
env = gym.make("SecurityLogStream-Hybrid-v0", db_path="exp_30d_heavy_v4.db")
obs, info = env.reset()
for _ in range(1000):
    obs, reward, terminated, truncated, info = env.step(env.action_space.sample())

Schema

All streams are SQLite databases (compressed with Zstandard) with a single events table:

CREATE TABLE events (
    id            INTEGER PRIMARY KEY AUTOINCREMENT,
    timestamp     TEXT NOT NULL,       -- ISO 8601 UTC
    source        TEXT NOT NULL,       -- event source (see below)
    raw_line      TEXT NOT NULL,       -- original log line or eBPF event text
    parsed        TEXT,                -- JSON parsed fields
    is_malicious  INTEGER,             -- 0=benign, 1=malicious
    campaign_id   TEXT,                -- attack campaign identifier
    attack_type   TEXT,                -- see attack types below
    attack_stage  TEXT,                -- MITRE ATT&CK stage
    severity      INTEGER,             -- 1-5
    session_id    TEXT,
    src_ip        TEXT,
    username      TEXT,
    service       TEXT
);

Event sources

auth_log — SSH authentication, PAM sessions
syslog — system daemon messages
web_access — nginx/apache access logs (Combined Log Format)
web_error — nginx/apache error logs
journal — systemd journal entries (JSON)
ebpf_process — kernel process exec/exit events (includes ppid + parent_comm)
ebpf_network — kernel connect/accept events (includes uid)
ebpf_file — kernel file open/unlink events

Attack types

Type	MITRE ATT&CK	Description
`brute_force`	T1110.001	SSH password brute force (paramiko)
`credential_stuffing`	T1110.004	SSH credential stuffing with unique credential pairs
`discovery`	T1046	SYN port scanning (scapy)
`web_exploit`	T1190	Log4Shell (CVE-2021-44228) JNDI injection; Redis Lua sandbox escape (CVE-2022-0543)
`execution`	T1059.004	Post-authentication Unix shell command execution

Composition methodology

Composed from benign_v4.db + campaigns_v2.db using StreamComposer with Poisson-scheduled attack injection and 24.2% eBPF downsampling (chosen to approximate a single busy host's kernel event rate). Attack campaigns follow per-stream MITRE ATT&CK-aligned distributions:

exp_7d_brute_v4: brute force only (1.0 campaigns/day)
exp_30d_heavy_v4: uniform 0.20 across discovery, brute_force, web_exploit, credential_stuffing, execution (10.0 campaigns/day)
exp_90d_v4 (ablation stream): discovery 0.35, brute_force 0.30, web_exploit 0.20, credential_stuffing 0.10, execution 0.05 (3.0 campaigns/day)
exp_365d_realistic_v4 (Zenodo only): discovery 0.30, brute_force 0.30, credential_stuffing 0.20, web_exploit 0.15, execution 0.05 (2.5 campaigns/day)

All streams pass nine validation checks via scripts/validate_labels.py (label consistency, raw-line spot checks, campaign boundaries, campaign type cross-validation, target-array consistency, attack-type distribution, temporal order, no unlabeled events, session coherence).

Labeling

Each attack campaign is executed against the target VM by the campaign orchestrator, which records start time, end time, and source IPs for every attack phase. The CampaignLabeler labels collected events using time-window + source-IP matching: an event is labeled malicious if its timestamp falls within a phase's time window AND its src_ip matches one of that phase's attacker IPs. Events with no src_ip (e.g., syslog daemon messages during an attack window) match on time window alone. eBPF kernel events use the identical labeler.

Malicious events receive five ground-truth fields: is_malicious=1, campaign_id, attack_type, attack_stage, and severity (1-5). Benign events have is_malicious=0 with the remaining fields NULL.

Intended use

Research on streaming continual reinforcement learning for cybersecurity defense
Online representation learning from raw text and structured kernel-event streams
Streaming deep-RL technique validation (LMS, ObGD, EMA normalization, sparse initialization)
Reward design in observation-suppressing environments
Comparative evaluation of detection approaches (learned vs. rule-based) on non-stationary data

Not intended for use

Production SOC tooling
A model of enterprise infrastructure
Zero-day generalization evaluation (all attack modules are known TTPs)
A substitute for real-world deployment validation
Multi-host lateral movement studies (single-target only)

The attack modules in this dataset target a purpose-built lab VM (isildur) and must not be directed at third-party infrastructure.

Provenance and PII

Benign traffic: 4 personal Linux servers running standard services. 7,915,858 log events. Hostnames, domains, and server IPs scrubbed via case-insensitive replacement to a single normalized target identity (isildur, 192.168.2.201).
Benign eBPF: 24-hour collections from 3 Debian 13 servers (hypervisor, public web server, GPU lab server). 3,243,383 kernel events.
Attack traffic: 10 scripted campaigns executed against a purpose-built Debian 11 VM with intentionally vulnerable services (OpenSSH, Log4j, Redis CVE-2022-0543). 60,468 events labeled via time-window + source-IP matching.
eBPF kernel events: Collected via BCC tracepoints (sys_enter_execve, sched_process_exit, sys_enter_connect, sys_enter_accept4, sys_enter_openat, sys_enter_unlinkat).
Hostile traffic filtering: Real-server benign baselines are filtered to remove attack traffic before inclusion. Network events from detected malicious IPs are excluded from the benign baseline to prevent contamination.
No third-party PII: All data originates from infrastructure owned or operated by the dataset author. No personal user data is included.

Citation

@dataset{lawson_security_gym_dataset_2026,
  author    = {Lawson, Keith},
  title     = {Security-Gym Dataset: Labeled Linux Log and eBPF
               Streams for Continual Learning Research},
  version   = {4.1},
  year      = {2026},
  doi       = {10.5281/zenodo.21763493},
  url       = {https://doi.org/10.5281/zenodo.21763493},
  license   = {Apache-2.0}
}

License

Apache-2.0. See LICENSE.

Downloads last month: 168

Total file size:

3.01 GB