Wi-Fi human sensing, boosted by latest progress in both system innovation and deep analytics, has demonstrated ever-increasing resolution of users' activities. Nonetheless, it may become a spy on users' private activities such as password entry or intimate social interactions. Existing countermeasures include signal obfuscation and adversarial perturbations to hamper and confuse Wi-Fi sensing, yet they both require substantial changes in Wi-Fi hardware/firmware, and they at most stay at user level in protection granularity. This paper presents Poison2Cure, the first semantic-level privacy-preserving framework for Wi-Fi human sensing systems, with full compatibility to any underlying hardware. The innovation behind Poison2Cure lies in feeding poisoned training data from (privacy-sensitive) users to the neural model for Wi-Fi sensing, degrading only the sensing for private activities while retaining that for regular ones. Moreover, we tackle the harsh conditions where the neural model is kept confidential and/or preceded by data cleansing. Our extensive evaluations demonstrate that Poison2Cure reduces over 76% of the accuracy for the private activities while keeping the accuracy for regular activities largely intact.