judgement-day-anon/Judgement-Day

Judgement Day — Review Subset

A 5,200-submission subset of the Judgement Day dataset, released for anonymous review of Judgement Day: An Anatomy of Successful Multimodal Attacks on Safety-Critical AI Systems. Each record is an attack input (audio, image, video, PDF, email, or text) that a participant submitted against a multimodal agent in one of eight safety-critical scenarios, and that caused at least one evaluated model to select an unsafe action.

Scenarios

id	scenario
track1_0	Emergency Triage: Multi-modal Resource Coordination
track1_1	Industrial Robot: Physical Sensory Deception
track1_2	Sports Integrity: Statistical Anomaly Normalization
track1_3	Dam Flood Control: Emergency Overtopping Deception
track2_0	Cybersecurity SOC: Incident Suppression
track2_1	Aircraft Predictive Maintenance: Hold Deferral
track2_2	Epidemiological Surveillance: Outbreak Alert Suppression
track2_3	Deepfake Detection: Forensic Finding Suppression

Layout

public_v1/
  index.csv        id, scenario_id, modality, path
  metadata.csv     same rows, HuggingFace file_name convention
  <scenario_id>/<modality>/<id>.<ext>

modality is one of audio, image, video, document, email, text. One file is one attack input, exactly as submitted. File ids are random UUIDs. Total size is about 14.5 GB.

Sampling

650 submissions per scenario, approximately balanced across the input channels available within each scenario, with priority given to attacks that breached multiple models. Breach rates on this subset are therefore higher than those of the full corpus.

modality	audio	document	image	video	text	email
count	2,489	908	641	529	328	305

What is not here

Per-model outcomes and strategy labels are withheld during review. The full dataset (all 50,136 successful attacks with per-model outcomes and strategy labels) will be released on publication. Scenario specifications and the evaluation code are in the paper's supplementary material.

License

Apache License 2.0.