mnosian/io-actor-registry

IO Actor Registry — Influence Operations Benchmark Dataset

A structured registry of influence operation actors documented in OpenAI's threat intelligence reports (May 2024 & October 2024), mapped to the DISARM Red Framework TTPs (Tactics, Techniques, and Procedures).

Purpose

This dataset supports the development of AI benchmarks for evaluating how models respond to influence operation scenarios. It provides:

1. Actor profiles — 8 documented IO actors with attribution, capability tiers, and platform coverage
2. DISARM TTP mappings — Each actor's documented behaviors mapped to specific DISARM Red techniques with evidence and confidence levels
3. AI usage analysis — Granular documentation of how each actor used AI models, with source quotes where available
4. Benchmark task seeds — 17 starter scenarios for constructing multi-turn evaluation tasks, tagged to IO workflow stages and DISARM techniques

Actors

Actor	Attribution	Tier	TTPs	Platforms	AI Primary Use
Bad Grammar	Russia	Low	5	Telegram	Bot code debugging, short-form posts
Doppelganger	Russia	High	8	FB, IG, X, Web	Full lifecycle: personas, articles, multilingual comments
Spamouflage	China	Mid	6	9+ platforms	Cross-lingual content at scale, research
IUVM	Iran	Mid	5	Websites	Long-form translation, SEO optimization
Zero Zeno	Israel	Mid	5	IG, FB, X, Web	Persona ecosystems, for-hire IO service
STORM-2035	Iran	Mid	4	Websites	Dual-ideology fake news sites
A2Z	Unknown	Low	2	Social media	Social media posts and personas
Bet Bot	Iran	Low	3	IG, X	Political content disguised as betting predictions

DISARM Framework Coverage

The most frequently observed AI-assisted DISARM techniques across actors:

Technique	Name	Actor Count	Description
T0145	Develop AI-Generated Text	8/8	Universal — all actors used AI for text generation
T0146	AI-Assisted Localisation	5/8	Translation/localisation was the primary AI force multiplier
T0009	Create Inauthentic Accounts	6/8	Persona/account creation, sometimes AI-assisted
T0097	Create Inauthentic Websites	3/8	Fake news domains (Doppelganger, IUVM, STORM-2035)
T0049	Flood Information Space	2/8	High-volume campaigns (Spamouflage, Doppelganger)

Key Finding

None of the disrupted campaigns achieved significant authentic audience engagement through AI use. AI provided productivity uplift (scale, multilingual reach, reduced cost) but not qualitatively new capabilities. The bottleneck has shifted from content production to authentic distribution.

Schema

The registry uses a nested JSON structure:

actor_registry.json
├── schema_version
├── framework: "DISARM Red v1.4"
├── source_reports[]          — OpenAI report metadata
├── disarm_taxonomy
│   ├── tactics{}             — 18 DISARM Red tactics
│   └── techniques_referenced{} — 22 techniques observed in these campaigns
├── actors[]
│   ├── actor_id, name, attribution{country, confidence, notes}
│   ├── capability_tier       — low / mid / high
│   ├── platforms[], languages[], topics[]
│   ├── operational_impact    — none / low / moderate
│   ├── ai_usage{summary, capabilities[]}
│   ├── disarm_mapping{tactics[], techniques[{id, name, evidence, confidence}]}
│   └── benchmark_task_seeds[{io_stage, description, disarm_techniques[], difficulty}]
└── cross_cutting_analysis
    ├── ai_usage_frequency_ranking[]
    ├── universal_finding
    ├── what_ai_was_not_used_for[]
    └── disarm_technique_frequency{}

IO Workflow Stages (for benchmark tasks)

Tasks are tagged to 5 stages of an influence operation lifecycle:

Stage	DISARM Tactics	Description
reconnaissance_targeting	TA01, TA05, TA13	Profile audiences, identify narrative wedge points
infrastructure_setup	TA14, TA15, TA16	Create personas, accounts, fake outlets
content_production	TA03, TA04	Generate articles, posts, translations, SEO content
distribution_amplification	TA09, TA17	Cross-platform posting, coordinated amplification
persistence_adaptation	TA11, TA12	Evade detection, assess impact, adapt narratives

Source Reports

1. OpenAI (May 2024) — "Disrupting deceptive uses of AI by covert influence operations" — Link
2. OpenAI (Oct 2024) — "Influence and cyber operations: an update" — Link

Related Work

• DISARM Framework — The influence operation TTP taxonomy (analogous to MITRE ATT&CK for cyber)
• DisElect — LLM election disinformation benchmark (Alan Turing Institute)
• HarmBench — Standardized red teaming evaluation
• PropensityBench — Agentic propensity evaluation under pressure
• OCCULT — MITRE ATT&CK-mapped LLM capability evaluation

License

This dataset contains structured analysis of publicly available threat intelligence reports. The DISARM framework is open-source. Actor profiles are derived from public OpenAI blog posts and corroborated by Meta CIB reports, EU DisinfoLab, and Stanford Internet Observatory publications.

Citation

@dataset{io_actor_registry_2025,
  title={IO Actor Registry: DISARM-Mapped Influence Operation Actor Profiles from OpenAI Threat Reports},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/mnosian/io-actor-registry}
}