Addiction is regarded as a disorder of inflexible choice with behavior dominated by immediate positive rewards over longer-term negative outcomes. However, the psychological mechanisms underlying the effects of self-administered drugs on behavioral flexibility are not well understood. To investigate whether drug exposure causes asymmetric effects on positive and negative outcomes we used a reversal learning procedure to assess how reward contingencies are utilized to guide behavior in rats previously exposed to intravenous cocaine self-administration (SA). Twenty-four rats were screened for anxiety in an open field prior to acquisition of cocaine SA over six daily sessions with subsequent long-access cocaine SA for 7 days. Control rats (n = 24) were trained to lever-press for food under a yoked schedule of reinforcement. Higher rates of cocaine SA were predicted by increased anxiety and preceded impaired reversal learning, expressed by a decrease in lose-shift as opposed to win-stay probability. A model-free reinforcement learning algorithm revealed that rats with high, but not low cocaine escalation failed to exploit previous reward learning and were more likely to repeat the same response as the previous trial. Eight-day withdrawal from high cocaine escalation was associated, respectively, with increased and decreased dopamine receptor D2 (DRD2) and serotonin receptor 2C (HTR2C) expression in the ventral striatum compared with controls. Dopamine receptor D1 (DRD1) expression was also significantly reduced in the orbitofrontal cortex of high cocaine-escalating rats. These findings indicate that withdrawal from escalated cocaine SA disrupts how negative feedback is used to guide goal-directed behavior for natural reinforcers and that trait anxiety may be a latent variable underlying this interaction.