rypow/dendroaspis-tetragon-hids

Dendroaspis Tetragon HIDS — Agent-Driven Attack Dataset

22 technique IDs (21 ATT&CK + 1 supply-chain capstone) · 19.3 M events · 6-day controlled collection window

A behavioral-event corpus collected from a single Linux LXC container under a controlled Atomic Red Team (ART) campaign, parsed and feature-engineered with a Tetragon-native pipeline. Released for:

• host-anomaly-detection research
• sequence modeling (Mamba, Transformers, etc.)
• ATT&CK-technique-detectability studies
• autonomous-attacker behavioral analysis — see below

The full provenance is in datacard.md and datasheet.md.

---

🤖 Agent-driven attack execution

The test set consists of events generated during a controlled penetration-testing campaign that was fully orchestrated by an AI agent.

The agent was responsible for:

• installing and configuring Atomic Red Team (ART)
• performing host-level reconnaissance
• selecting applicable ATT&CK techniques
• constructing the attack campaign
• executing the techniques within the target environment

This is a fundamental shift from traditional HIDS datasets:

Dataset type	Execution model
Traditional HIDS corpora	Human-scripted attacks
This dataset	Agent-driven attack orchestration

Operational model

The dataset captures an end-to-end autonomous attack loop:

Observation → Reconnaissance → Planning → Tooling → Execution → Iteration

The resulting telemetry reflects not only attack techniques, but the behavioral footprint of an autonomous attacker operating in a real system.

Why this matters

Agent-driven attacks introduce:

• structured but adaptive execution patterns
• tool installation and environment-preparation behavior
• continuous recon → execution feedback loops
• multi-stage dependency chains across techniques

This enables research into:

• detection of autonomous-attacker behavior
• distinguishing human vs agent execution patterns
• modeling attacks as sequences of intent
• identifying planning and coordination signals in telemetry

---

Capstone: agent-driven supply-chain attack (CHAIN.CAP01)

In addition to 21 individual ATT&CK techniques, the dataset includes a capstone scenario simulating a malicious npm supply-chain compromise, inspired by real-world self-propagating npm worms (e.g., the 2025 Shai-Hulud incident).

In this scenario, the agent:

• installs a dependency via npm install
• triggers execution through package lifecycle scripts (e.g. postinstall)
• performs local reconnaissance and file access
• executes follow-on actions including potential credential access and exfiltration

This produces a multi-stage attack chain compressed into a short execution window (≈14 seconds, 34,130 events), labeled CHAIN.CAP01. Unlike the individual ATT&CK techniques, the capstone is a coordinated attack narrative combining:

• execution
• discovery
• credential access
• defense evasion
• exfiltration behaviors

within a single autonomous workflow.

Why the capstone matters

This scenario reflects a realistic modern attack vector:

• software supply-chain compromise
• developer-environment targeting
• automated post-install exploitation

and offers a unique opportunity to study:

• cross-technique behavioral chaining
• short-duration, high-density attack bursts
• agent-driven multi-stage execution patterns

---

Splits at a glance

Split	Rows	Collection window (UTC)†	Attack rate
train	16,951,950	2026-04-14 01:41 → 2026-04-19 22:09	0.0 %
test	2,349,965	2026-04-19 21:46 → 2026-04-20 02:01	30.0 %

All ART attack intervals fall in the test split. Train is benign-only; unsupervised / self-supervised methods are the natural fit.

† Pre-window event timestamps. ≈295 events (267 train + 28 test, 0.0015 % of the corpus) carry event_time values that predate the stated collection window. All have f_is_procfs_walk = 1: parser-synthesized process_exec records for daemons (systemd, bash, dbus-daemon, polkitd, cron, sshd, etc.) that were already running when Tetragon collection started, using the daemon's original process_start_time as the event timestamp. They preserve lineage roots for downstream events. Filter WHERE f_is_procfs_walk = 0 to exclude them from time-window analysis.

---

Top techniques by event volume (test split)

ID	Technique	Test events
T1059.004	Unix Shell	175,797
T1083	File and Directory Discovery	154,581
T1552.001	Credentials In Files	97,493
T1543.002	Systemd Service	47,901
CHAIN.CAP01	Capstone Chain v0.1 (Supply-Chain Narrative)	34,130
T1548.003	Sudo and Sudo Caching	24,229
T1059.006	Python	22,675
T1562.001	Disable or Modify Tools	18,610
T1195.002	Compromise Software Supply Chain	15,703
T1548.001	Setuid and Setgid	12,461

Full per-technique table (22 IDs across 8 ATT&CK tactics) lives in labels.csv and datacard.md.

---

Headline numbers (Run 2, post-rebuild)

Scorer	Test AUROC
Mamba MEM-FA (field-aware)	0.7513
Mamba NLL (autoregressive)	0.7411
IsolationForest (best classical)	0.6658
n-gram (token-level)	0.5672
XGBoost	0.5000

Mamba's sequence stack adds ≈0.085 AUROC over the strongest classical baseline. See datacard.md for the per-technique breakdown and the per-window inversion analysis on unsupervised classicals.

---

Quick start

from datasets import load_dataset
import pandas as pd

ds = load_dataset("rypow/dendroaspis-tetragon-hids", streaming=True)

event = next(iter(ds["train"]))
print(event["f_event_type"], event["f_action_family"], event["token"])

labels = pd.read_csv("labels.csv", parse_dates=["start_ts", "end_ts"])
labels["start_ts"] = labels["start_ts"].dt.tz_convert("UTC")
labels["end_ts"]   = labels["end_ts"].dt.tz_convert("UTC")
print(f"{len(labels)} intervals · {labels['technique_id'].nunique()} techniques")

To interpret the integer-coded categorical columns, load reproduction/feature_definitions.json — every code (e.g. f_action_family=6) maps to a label and a one-line semantic description.

---

Anonymization

The released parquet has been fully anonymized via a component-first deep aliasing pass:

• Basenames (proc_binary_basename, parent_binary_basename, root_ancestor_basename): an allowlist of distro / dev-tool basenames passes through verbatim (bash, node, python3.12, claude, gh, npm, code, git, …); custom binaries become custom_tool_NNN; VSCode commit-hash binaries collapse to vscode_server_bin.
• token and parent_child_pair are regenerated from the aliased basenames via the canonical format_token (so the original basename never re-enters these composite columns).
• proc_cwd_normalized is path-rewritten (/home/<user> → /home/_user_, /mnt/projects/<X> → /workspace/_project_, etc.).
• proc_exec_id and process_tree_root_exec_id are pseudonymized via salted HMAC-SHA256 to exec_<32hex>. Equality joins are preserved (lineage walks still work).
• A defensive global string sweep substitutes rypow → research_user, prox1 → sandbox_host, dendroaspis → agent_platform_v02, ryan_admin → privileged_user across every string column.

Post-write assertions confirm zero PII hits across raw and base64-decoded forms of all 12 string columns. Full alias map at reproduction/alias_map.json; the realised PASS set at reproduction/allowlist_basenames.txt.

See datacard.md "Personal and sensitive information" for the per-column decision log.

---

Citation

@dataset{dendroaspis_v0_2_2026,
  title   = {Dendroaspis Tetragon HIDS Dataset (Agent-Driven Attack Variant)},
  author  = {Powers, Ryan William},
  year    = {2026},
  version = {0.2.1}
}

The canonical BibTeX is also at dendroaspis.bib; update the url and doi fields once those are minted.

---

Pointers

• Full feature documentation: datacard.md
• Provenance + collection process: datasheet.md
• Reproducing collection: reproduction/REPRODUCE.md
• Integer-code → label tables: reproduction/feature_definitions.json
• Anonymizer alias map: reproduction/alias_map.json
• Per-file integrity: SHA256SUMS, machine manifest at RELEASE.json
• Github: https://github.com/ryypow/dendroaspis

---

License

CC BY 4.0 — see LICENSE. Code under reproduction/scripts/ (in the project repository) is MIT-licensed.

Intended use + dual-use disclaimer

Defensive research only. The attack telemetry is real ATT&CK-technique execution but generated in an isolated lab LXC container; the corpus is not an attack template, not a payload library, and not a substitute for production EDR data. Operators who reuse the bundled tracing policies + parser to collect their own corpus are responsible for their own sandboxing.

See datacard.md "Intended use + dual-use disclaimer" for the full statement.