goldenowlosint/iranian-regime-influence-machinery-against-israeli-society

Exposing the Iranian Regime’s Influence Machinery Against Israeli Society — Account Dataset

Companion dataset to the Golden Owl research Exposing the Iranian Regime’s Influence Machinery Against Israeli Society. Read the full report (English and Spanish) on the Golden Owl website.

An account-level dataset of 694 X (Twitter) accounts assessed as part of, or connected to, a coordinated foreign information manipulation and interference (FIMI) operation targeting Israeli public discourse ahead of the 27 October 2026 Knesset election. Each account is listed with its confidence tier, the indicators that support the assessment, X's own transparency labels and its links to the rest of the network.

The dataset accompanies the Golden Owl report Exposing the Iranian Regime’s Influence Machinery Against Israeli Society (research period 11 August – 11 October 2026; snapshot 11 October 2026). It extends the network documented in September in The Other Face of War: Iranian Regime–Linked Manipulation of Israel's 2026 Election Debate.

About the research

This dataset is the account-level evidence base of Exposing the Iranian Regime’s Influence Machinery Against Israeli Society, an open-source investigation by Golden Owl published in October 2026.

The research documents an Iranian regime-operated apparatus of Hebrew-language personas that speaks to Israelis as Israelis on X in the final weeks before the 27 October 2026 Knesset election. It extends the network Golden Owl exposed in September and maps it in full:

• Scale: 560,439 data points processed; a coordinated network of at least 694 accounts, 445 of them attributed to the Iranian-linked operation with high or medium-high confidence, and 623 outside the network documented in September.
• Construction: fabricated Israeli identities (Hebrew names, Israeli cities, flags and biographies), accounts created in coordinated batches during the June 2025 war, identities held in reserve as several accounts each, a shared text library and Persian-keyboard traces inside the Hebrew.
• Reach: collected network posts drew at least 4.7 million views and 160,663 interactions between 11 August and 11 October 2026, largely through reply seeding under politicians, journalists and popular accounts.
• Objective: not support for one party but the erosion of Israeli trust in institutions, the military and the electoral process, through narratives of betrayal, corruption, collapse and emigration.

What you see in this dataset. Each row is one X account from that investigation: its confidence tier (T1 to T4), the indicators behind the assessment, X’s own transparency labels (location and app build), its creation date and audience, and its links to the rest of the network (accounts it retweets or follows, and accounts it shares text with). The report explains the findings, the evidence boards and the method; this dataset lets readers verify each account on its own evidence.

Key figures

Measure	Value
Accounts listed	694
T1 — high confidence (attributed)	336
T2 — medium-high confidence (attributed)	109
T3 — medium confidence (investigative lead)	179
T4 — under monitoring (investigative lead)	70
Accounts attributed to the Iranian-linked operation (T1 + T2)	445
Accounts outside the network documented in September	623

Only T1 and T2 accounts are attributed by Golden Owl to the Iranian-linked operation. T3 and T4 accounts are investigative leads and are not presented as established participants.

Files

File	Format
golden_owl_october2026_accounts.jsonl	JSON Lines, one account per line, typed values (numbers, booleans, and lists for indicators and the network-link fields). Used by the dataset viewer
golden_owl_october2026_accounts.csv	The same 694 rows as flat CSV; list fields are joined with ; (indicators) or , (accounts)

Fields

Column	Description
tier	Confidence tier: T1, T2, T3 or T4
confidence	high, medium-high, medium or under monitoring
score	Sum of indicator weights. Informative only: tiers follow combination rules across evidence families, not the score
cluster	Analytical cluster the account belongs to (for example "Iran-operated Israeli personas (new)")
handle	X username at collection time (handles change; use user_id)
user_id	Stable X account ID
display_name	Display name at collection time
declared_location	Location the account declares on its profile
x_based_in	X's own "Account based in" label
x_connected_via	X's own "Connected via" label (app build, for example "Iran Android App")
x_location_flagged	Whether X flags its own location label as possibly inaccurate
username_changes	Number of username changes recorded by X
created_at	Account creation date
followers, following	Follower and following counts at collection time
posts_lifetime	Lifetime post count shown on the profile, where available
israel_facing	Whether the account presents as Israeli or writes for an Israeli audience in Hebrew
indicators	Indicator codes supporting the assessment (dictionary below)
retweeted_network_accounts	Network accounts it retweeted during the research period
follows_network_accounts	Network accounts it follows
shared_text_partners	Accounts with which it shares verbatim or near-verbatim text
label_assessment	Content-label assessment of the account (for example likely_inauthentic_persona)
label_rationale	Short rationale for the content label
tier_adjustment	Set when the rule-based tier was capped or set on analyst review
profile_url	Link to the public X profile

Indicator families

Accounts were assessed with 54 indicators in five evidence families plus two zero-weight information notes:

• P — platform-attested origin: X's location and app-build labels (Iranian app build, Israel-facing Hebrew account located in the Arab world, Turkey or a country without an Israeli diaspora, third-country app builds).
• I — identity and history: creation date, username changes, auto-generated handles, copies of another persona's identity, look-alike characters in the declared location.
• B — behaviour: amplification of several network accounts, high-volume retweeting, self-retweet bursts, reply bursts, retweets synchronised within 60 seconds.
• C — content: text shared with personas, the June 2025 wartime hashtag kit, Arabic or Persian characters inside Hebrew, machine-assisted Hebrew, non-native Hebrew and gender inconsistencies.
• N — network position: follow ties to Iran-labelled personas, links to Iranian Hebrew-language outlets, previous classification.
• X — information notes (weight 0).

The full dictionary with weights is published in the report (Appendix D).

Usage

from datasets import load_dataset

ds = load_dataset("goldenowlosint/iranian-regime-influence-machinery-against-israeli-society", split="train")
attributed = ds.filter(lambda r: r["tier"] in ("T1", "T2"))

import pandas as pd

df = pd.read_json("hf://datasets/goldenowlosint/iranian-regime-influence-machinery-against-israeli-society/golden_owl_october2026_accounts.jsonl", lines=True)
df.groupby("tier").size()

Methodology summary

A mixed-methods open-source intelligence approach combining automated processing with human analytical validation. Golden Owl's proprietary intelligence technology processed 560,439 distinct data points from X between 11 August and 11 October 2026 (account profiles, posts, relationships, platform transparency records and amplification activity), building outward from the network identified in September. Accounts were examined through natural language processing, social network analysis, behavioural analysis, identity and platform analysis, and cross-source correlation. Tiers were assigned through predefined rules combining indicators across evidence families, supported by weighted scores and analyst validation; no single indicator is sufficient for a high tier.

Limitations & caveats

• Platform location and app-build labels are X's own inference and can be affected by VPNs, travel or platform errors; they are always combined with other indicators.
• Collection does not represent the complete history of every account. Counts are as recorded on 11 October 2026.
• Content labels and their rationales are machine-assisted and should be read together with the indicators.
• Accounts can be suspended, renamed, sold or hijacked after collection. Re-check before acting.
• Genuine users whose only activity was following or amplifying network content are not listed.

Ethics & responsible use

Handles identify public online accounts. They do not establish the legal identity of the persons controlling them. Inclusion is an analytical assessment of observable conduct — coordination, deceptive identity presentation and platform-attested foreign operation — not an allegation that any identifiable person committed an offence or knowingly acted for a foreign state. Political opinion was never used as an indicator. Cite the tier and the indicators, not just the handle, and do not use this dataset to harass or dox anyone.

Citation

Please cite the research report together with the dataset: Golden Owl (2026), Exposing the Iranian Regime’s Influence Machinery Against Israeli Society, https://goldenowl.ai/resources/research/Exposing-the-Iranian-Regime-Influence-Machinery-Against-Israeli-Society

@dataset{goldenowl_iranian_regime_influence_machinery_2026,
  author    = {{Golden Owl}},
  title     = {Exposing the Iranian Regime’s Influence Machinery Against Israeli Society — Account Dataset},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/goldenowlosint/iranian-regime-influence-machinery-against-israeli-society}
}

Collaboration

This research was independently initiated and conducted by Golden Owl®. Golden Owl® is collaborating with YSocial on further investigations into the identified accounts and networks.

License

Released under CC BY 4.0. Account handles, display names and profile fields are reproduced from public X profiles for research and accountability purposes.