Many studies have explored the TCP throughput problem in DiffServ networks. Several new marking schemes have been proposed in order to solve this problem. Even if these marking schemes give good results in the case of per-flow conditioning, they need complex measurements. In this paper we propose a Penalty Shaper (PS) which is able to profile a set of TCP flows so as to improve its conformance to a desired target. The main novelty of this shaper is that the shaping applies a penalty delay which depends on the out-profile losses in a DiffServ network. This penalty shaping can be used with any classic conditioner such as a token bucket marker (TBM) or a time sliding window marker (TSWM). We made an evaluation of the Penalty Shaper on a real testbed and showed that the proposed scheme is easily deployable and allows for a set of TCP flows to achieve its target rate.