AI safety data

Real-world data for AI safety.

We license safety data drawn from real usage of ablated open-weight models. It spans cybersecurity, NSFW content, agentic tool use and other sensitive domains. None of it is synthetic.

Data categories

Each category contains real requests and the model's actual responses, including requests that safety-trained models refuse.

Cybersecurity

Offensive and defensive security work, including reconnaissance, vulnerability research, exploitation tooling and penetration testing.

NSFW content

Explicit adult content and roleplay, for training and evaluating content classifiers and policy boundaries.

Agentic trajectories

Multi-step agent runs with tool calls and reasoning, including safety-relevant actions that occur partway through a run.

Refusal boundaries

Requests that safety-trained models decline, covering both harmful requests and legitimate requests that are refused in error.

Dual-use automation

Scraping, anti-bot evasion, account automation and similar workflows.

Additional domains

Further sensitive categories and non-English data are available on request.

Real usage, not synthetic traces

Our data is sourced from our proprietary inference providers, which serve ablated open-weight models to real users. Because these models do not refuse, the data records what users actually ask for and what models actually do.

Real users

Every record originates from an actual user request in production. There is no synthetic generation or simulated red teaming.

Ablated models

Refusal behavior is removed, so interactions continue past the point where safety-trained models stop.

Complete records

Prompts, responses, reasoning and tool calls are retained in full, with multi-turn sessions kept intact.

Data handling

Every release follows the same standards.

Consent

Data is licensed only where users have agreed to its use.

De-identification

Personal data, credentials and identifying details are removed before delivery.

Prohibited content

Content involving minors is excluded and reported as required by law.

Licensing

Licensed to vetted organizations for safety research and development, under NDA and defined terms of use.

Request a sample

Samples are available under NDA to organizations working on AI safety. Contact us to discuss available data and licensing terms.

Contact us
data@dreamingsand.com