synthitd — Multi-Domain Insider-Threat Benchmark

Synthetic, fully reproducible, multi-surface workplace telemetry for insider-threat detection (ITD), with graded ground-truth labels grown from a Critical-Pathway intent model rather than injected as scripted attacks. This release contains three data-sensitive organisations — technology, finance, healthcare — each 500 employees over 120 days. Pick a domain with the dataset viewer's config selector (default: tech).

Defensive / academic use only. Behaviour is modelled as detection-relevant observables and MITRE ATT&CK technique labels; no operational attack instructions. Fully synthetic — no real people, accounts, credentials, or messages.

Code, generator, benchmark, paper: https://github.com/AlaouiMdaghriAhmed/LLM_DataGen

Instances in this release

domain	employees	activated/insiders	anomalous-benign	events (full)	events (observed)	precursor	malicious
tech	500	11/15	40	5,696,495	5,001,816	362	133
finance	500	11/15	40	5,629,758	5,019,852	399	92
healthcare	500	13/15	40	5,658,779	5,081,705	439	189

Layout (per domain <d>/)

file	description
<d>/labels_userday.csv	viewer config — per (user, day): graded label, latent risk, counts
<d>/events.jsonl	full labelled ground-truth stream (one JSON event per line)
<d>/observed.jsonl	stream after the sensor-degradation filter (what a detector sees)
<d>/employees.json	org: roles, access, OCEAN traits, predisposition, population
<d>/episodes.json	one record per activated insider episode (pathway, onset, MITRE techniques)
<d>/ground_truth.json	per-employee latent risk-trajectory summary
<d>/manifest.json	config + summary + observability reports
<d>/cert/	CERT r6.2-compatible export (+ cert/extended/ modern surfaces)

Event schema (events.jsonl)

event_id, ts, channel, actor, action, attrs, label, episode_id, technique_ids, risk_state, render, causal_parent. Channels: logon, idp, device, file, email, http, chat, ticket, vcs, dlp, endpoint, hr.

Labels (graded)

benign · anomalous_benign (hard negative, paired with an HR-context event) · precursor (concerning activity on an armed pathway, before the act) · malicious.

Reproduce

Every instance is a pure function of its config and regenerates byte-for-byte:

pip install -e ".[dev]"
python scripts/build_release.py 500 120 out/synthitd_release

Citation

@misc{synthitd2026,
  title        = {synthitd: A Reproducible, Endogenous-Intent Synthetic Benchmark for Insider-Threat Detection},
  author       = {Alaoui Mdaghri, Ahmed},
  year         = {2026},
  howpublished = {\url{https://github.com/AlaouiMdaghriAhmed/LLM_DataGen}}
}