aeriesec/orgforge-insider-threat

OrgForge Insider Threat Detection Benchmark

Structured security telemetry for benchmarking LLM-based insider threat detection. No embedder required — the corpus is pre-structured JSONL/Parquet. Ground truth is derived deterministically from the simulation's event log.

Dataset Summary

This dataset was produced by OrgForge, an event-driven organisation simulator, with the insider threat module enabled. A corporate simulation runs for 51 days; configured threat subjects exhibit realistic anomalous behaviors across multiple artifact surfaces. Detection agents read the observable telemetry stream and must correlate signals across days and record types without access to ground truth labels.

Property	Value
Company	Apex Athletics
Industry	sports technology
Simulation days	51
Threat subjects	3
DLP noise ratio	0.4
Log format	jsonl
Total observable records	2,904
True positive records	106
False positive (noise) records	2,798
Baseline (clean) records	3,930
Best leaderboard verdict F1	0.8 (us.anthropic.claude-opus-4-6-v1)

Threat Subjects

Name	Class	Onset day	Behaviors
Jordan	negligent	5	secret_in_commit
Tasha	disgruntled	10	sentiment_drift, cross_dept_snooping, unusual_hours_access
Jax	malicious	18	data_exfil_email, excessive_repo_cloning, unusual_hours_access, sentiment_drift, host_data_hoarding, social_engineering

Subjects behave normally before their onset day. Pre-onset records are clean true negatives and form the detection baseline.

Telemetry Files

All files are in telemetry/. The observable stream contains no ground truth labels — true_positive, threat_class, and behavior fields appear only in ground_truth/.

File	Description	Rows
access_log-00000.parquet	Full observable stream	2,904
baseline_telemetry-00000.parquet	Pre-onset clean records only	3,930
idp_auth-00000.parquet	IDP authentication events	2,831
host_events-00000.parquet	Host-level staging events	6

Record types in the observable stream

| idp_auth | 2,831 | | slack_message | 30 | | repo_access | 29 | | email_send | 7 | | host_event | 6 | | phone_call | 1 |

Telemetry schema

Core fields present on every record:

Column	Type	Description
record_type	str	Event category (see table above)
day	int	Simulation day (1-indexed)
date	str	ISO date
timestamp	str	ISO datetime (SimClock-accurate)
actor	str	Employee name — no threat annotation
extra	str	JSON string of additional detail fields

Selected promoted detail columns (present when applicable):

Column	Type	Description
auth_result	str	IDP: success or mfa_failure
src_ip	str	IDP: source IP address
new_device	bool	IDP: device not in employee's known profile
anomalous_ip	bool	IDP: IP outside corporate range
ghost_login	bool	IDP: disgruntled ghost login pattern
preceded_by_call_record	bool	IDP: auth preceded by phone_call (vishing)
action	str	Host: bulk_file_copy, archive_creation, archive_move
hoarding_trail_start_day	int	Host: links phase 3 back to phase 1
destination_type	str	Host: cloud_sync_dir or removable_media
outside_business_hours	bool	All: access outside 09:00–18:00
pattern	str	Social engineering: spear_phishing, slack_pretexting, vishing_breadcrumb, trust_building
preceded_by_call_record	bool	IDP: marks vishing post-auth events

Ground Truth

ground_truth/ground_truth-00000.parquet contains all records from the observable stream with three additional columns:

Column	Type	Description
true_positive	bool	True if this record is a genuine threat signal
threat_class	str	negligent, disgruntled, malicious, or null
behavior	str	Behavior name, or null for false positive records

This file is held out during evaluation. Detection agents read only the telemetry files. Ground truth is used for scoring only.

Behaviors present in ground truth

| unusual_hours_access | 62 | | sentiment_drift | 27 | | data_exfil_email | 7 | | host_data_hoarding | 6 | | social_engineering | 4 |

Threat class distribution

| malicious | 61 | | disgruntled | 45 |

Detection Scenarios

The dataset is designed to benchmark reasoning under noise, not keyword matching. Key scenarios:

Ghost login (disgruntled): idp_auth success present but no Slack, Jira, or email artifacts follow in the same time window. corroborating_activity_expected: false marks these records.

Vishing (malicious): A phone_call record is followed within minutes by an idp_auth success filed under the target's actor name, not the attacker's. preceded_by_call_record: true and call_to_auth_gap_minutes link them. Per-actor triage agents miss this entirely — cross-actor correlation is required.

Host data hoarding (malicious/disgruntled): Three-phase trail spread across consecutive days: bulk_file_copy → archive_creation → archive_move. The hoarding_trail_start_day field in phase 3 records links back to phase 1. Single-day triage agents will always miss the complete trail.

Trust building (social engineering): A benign inbound email with sender_in_known_contacts: false precedes a follow-up attack 3–5 days later. The first contact is a clean false negative in isolation.

Leaderboard

leaderboard/insider_threat_leaderboard.csv contains a frozen snapshot of all model runs against this export. The full JSON version is also included.

Columns:

Column	Description	Better
triage_f1	F1 on escalation decisions	↑
verdict_f1	F1 on full case verdicts	↑
baseline_fp_rate	FP rate on clean baseline period	↓
onset_sensitivity	Fraction of pre-onset escalations	↓
vishing_detected	Did the agent correlate phone_call → idp_auth?	✓
host_trail_reconstructed	Did the agent cite all 3 hoarding phases?	✓

To add a row, run eval_insider_threat.py against this export and append the output to these files.

Evaluation Pipeline

# Build baseline (pre-onset clean records)
python build_baseline_telemetry.py --export-dir ./export

# Run detection pipeline (one command per model)
python eval_insider_threat.py \
    --model anthropic.claude-opus-4-5-20251101-v1:0 \
    --export-dir ./export

# Launch leaderboard UI
python app.py   # (insider threat Gradio app)

No embedder, no MongoDB, no vector database required. Credentials: AWS Bedrock (standard credential chain).

Citation

@misc{orgforge_it2026,
  title  = {OrgForge Insider Threat Detection Benchmark},
  author = {Jeffrey Flynt},
  year   = {2026},
  note   = {Synthetic benchmark generated by the OrgForge insider threat simulator}
}

Related Paper

https://arxiv.org/abs/2603.14997

License

MIT. The simulation engine that produced this dataset is independently licensed; see the OrgForge repository for details.