The best models have seen the worst
Get custom and off-the-shelf datasets built from real-world adversarial intelligence curated by experts for SFT, DPO, and classifier development.
Rare Adversarial Intelligence
Train models with rare abuse and attack patterns curated by our experts from real-world threats, including the darknet and closed communities.
150+ Safety and Security Experts
Address risks that require specialized knowledge with datasets curated and reviewed by in-house safety and security experts.
Localized Data, Not Generic Translation
Improve recognition of harmful behavior across languages with data curated by native-language experts who understand cultural context.
Delivered to
Your Criteria
Get existing or custom datasets tailored to your policies and model gaps, with every delivery validated against your acceptance criteria.
Adversarial Training Data for Safe, Secure, and Capable Models
Indirect prompt injection attacks created to expand attack variety beyond typical industry datasets.
~250 attacks
Text
In-house GenAI security researchers
Expanding indirect prompt injection coverage
Under two weeks
~50% attack success rate
Safe and unsafe responses to questions involving chemical, biological, radiological, and nuclear risks.
Pathogens, outbreaks, biosecurity, and public health
CBRN expert network
Responses tailored to different actors and intents
Multimodal violative content sourced and curated for a top-three foundation model organization.
Thousands of images and videos
Image and video
22
Bad-actor communities
Diversity and fit with customer requirements
Under three weeks
Testing enterprise agents in simulated workflows with hidden risks.
Hundreds of scenarios
Biased interviewing, off-topic behavior, and other risks
Simulated email and messaging platforms
Responsible AI training data
Leading agentic model
Multilingual prompt-response pairs recorded for training a leading frontier model.
Thousands of prompt-response pairs
Audio
20
Pitch, accent, dialect, and gender diversity
Frontier-model training
Under two weeks
We've Been Down The Rabbit Hole
The Rabbit Hole is Alice’s database of evil built over a decade across hundreds of languages and cultures. Our researchers turn these patterns into hidden attacks that test whether agents complete their tasks while resisting manipulation.
Talk to Our Experts“Alice supports our most complex and high priority red teaming needs. We like working with them because they are leaders in this space, understand the shifting adversarial landscape, and the output is always high quality. They are great partners who quickly adapt to our ever changing needs.”
Train on the Threats Your Models Will Face
Get custom or off-the-shelf datasets for post-training and classifier development, grounded in real-world adversarial patterns and expert knowledge.