Type: url
Original: https://www.anthropic.com/news/building-safeguards-for-claude
Status: generated
Summary
Anthropic’s account of how its Safeguards team builds protections across a model’s lifecycle: policy development guided by a Unified Harm Framework and external Policy Vulnerability Testing, collaboration with fine-tuning teams, pre-release safety, risk and bias evaluations reported in system cards, real-time detection and enforcement, and ongoing monitoring. Its runtime layer is built on classifiers — prompted or specially fine-tuned Claude models detecting policy violations in real time, several deployable at once — which can trigger response steering that adds instructions to a live request’s system prompt or stops a response outright, alongside account-level enforcement. It also records that pre-launch evaluation of the computer use tool found the capability could augment spam generation and distribution, leading to new detection methods, account-level tool disabling and prompt-injection protections before launch.