Humans are more likely to help those who they have observed helping others previously. Individuals may thus benefit from being altruistic without direct reciprocity of recipients but due to gains in 'image' and associated indirect reciprocity. I suggest, however, that image-scoring individuals may be exploitable by cheaters if pay-offs vary between interactions. I illustrate this point with data on cleaner-client reef fish interactions. I show the following: (i) there is strong variation between cleaners with respect to cheating of clients (i.e. feeding on client tissue instead of parasites); (ii) clients approach cleaners, that they observe cooperating with their current client and avoid cleaners that they observe cheating; (iii) cleaners that cheat frequently are avoided more frequently than more cooperative cleaners (iv) cleaners that cheat frequently behave altruistically towards their smallest client species; (v) altruistic acts are followed by exploitative interactions. Thus, it appears that cleaners indeed have an image score, which selects for cooperative cleaners. However, cheating cleaners use altruism in potentially low-pay-off interactions to deceive and attract image-scoring clients that will be exploited.