The integration of sensing, communications, array signal processing, etc. into 6G mobile networks has ushered in an era of heightened situational awareness. However, this progress brings forth significant concerns regarding privacy and security, particularly due to the proliferation of devices equipped with radar-like sensing capability, including malicious ones. In response, this paper proposes a novel actor-critic (AC) method-based frequency selection scheme for noise jamming, in order to effectively counter malicious multifunction frequency agility sensing. In the meanwhile, to mitigate potential interference (caused by sidelobes of the jamming beam) with uplink transmissions conducted by legitimate but non-cooperative users, a robust action correction mechanism, which is capable of learning and predicting the spectrum utilization state, is proposed to find feasible but near-optimal frequency configuration for jamming. Numerical results demonstrate that benefiting from the robust action correction mechanism, the proposed AC-based safe jamming can not only make the malicious sensing device continuously get stuck in the searching mode but also guarantee minimal disruption to the legitimate non-cooperative users.