Regular expressions are ubiquitous. They are used for diverse purposes, including input validation and firewalls. Unfortunately, they can also lead to a security vulnerability called ReDoS(Regular Expression Denial of Service), caused by a super-linear worst-case execution time during regex matching. ReDoS has a serious and wide impact: since applications written in most programming languages can be vulnerable to it, ReDoS has caused outages at prominent web services including Cloudflare and Stack Overflow. Due to the severity and prevalence of ReDoS, past work proposed mechanisms to identify and repair regexes. In this work, we set a different goal: helping developers avoid introducing regexes that could trigger ReDoS in the first place. A necessary condition for a regex to trigger ReDoS is to be infinitely ambiguous (IA). We propose a theory and a collection of anti-patterns to characterize infinitely ambiguous (IA) regexes. We evaluate our proposed anti-patterns in two complementary ways: quantitatively, over a dataset of 209,188 regexes from open- source software; and qualitatively, by observing humans using them in practice. In our large-scale evaluation, our anti-patterns characterized IA regexes with 100% precision and 99% recall, showing that they can capture the large majority of IA regexes, even when they are a simplified version of our theory. In our human experiment, practitioners applying our anti-patterns correctly assessed whether the regex that they were composing was IA or not in all of our studied regex-composition tasks.