Open-weight models cannot be considered safe because their safeguards can be easily removed

The field of preventing harm from advanced AI systems, spanning alignment, evaluation, and decisions to slow or withhold models whose capabilities outpace their safeguards.

Systems trained on data to learn patterns and make predictions or generate outputs, such as text, images, or decisions, by approximating relationships between inputs and outputs without being explicitly programmed for each task.

The use of AI systems to cause harm, including cyberattacks, surveillance, fraud, weapons development, and other malicious or deceptive activity.

AI models whose trained weights are published for anyone to download, run, and modify.

Moonshot AI's 3T-class model with a 1-million-token context window, released on October 23, 2026 through Moonshot's apps and API, with full open weights promised by November 4, 2026. Moonshot says it removed cyberattack-related training data before release, cutting the model's internal exploit-test score from 90.5% to 81%.