dual-use
Knowledge or capability that helps beneficial work and harmful work equally well, so a tool cannot simply block the bad half.
Dual-use describes something whose good and bad applications are the same thing viewed from different angles. The classic example is biology: developing a live vaccine means growing the pathogen you are trying to prevent, and the blood pressure drug captopril came from isolating the toxic part of snake venom. Chemistry, cybersecurity and genetics all have the same property.
For AI companies this is the hardest safety problem there is, because you cannot filter for a topic, only for intent, and intent is invisible. A researcher and a bad actor can ask word-for-word identical questions. That is why models are often blocked from whole areas rather than specific requests, and why the blocking is usually too broad at first and gets refined afterward.