ConceptGuardrails / 安全护栏 / 模型护栏
Capability Guardrails
Technical and policy-level constraint mechanisms implemented by AI vendors to restrict models from generating harmful outputs. These typically include RLHF-based value alignment, inference-layer content filtering, and usage policy and legal constraints, designed to prevent models from being exploited for attacks or harmful purposes.
Timeline (last 90 days)
Sep 18
Capability guardrails typically include value alignment during RLHF training, content filtering at the inference layer, and legal constraints from usage policies.
Unverified50%