Constructing collision-resistant hash families (CRHFs) from one-way functions is a long-standing open problem and source of frustration in theoretical cryptography. In fact, there are strong negative results: black-box separations from one-way functions that are $2^{-(1-o(1))n}$-secure against polynomial time adversaries (Simon, EUROCRYPT ’98) and even from indistinguishability obfuscation (Asharov and Segev, FOCS ’15). In this work, we formulate a mild strengthening of exponentially secure one-way functions, and we construct CRHFs from such functions. Specifically, our security notion requires that every polynomial time algorithm has at most $2^{-n – \omega(\log(n))}$ probability of inverting two independent challenges. More generally, we consider the problem of simultaneously inverting $k$ functions $f_1,\ldots, f_k$, which we say constitute a “one-way product function” (OWPF). We show that sufficiently hard OWPFs yield hash families that are multi-input correlation intractable (Canetti, Goldreich, and Halevi, STOC ’98) with respect to all sparse (bounded arity) output relations. Additionally assuming indistinguishability obfuscation, we construct hash families that achieve a broader notion of correlation intractability, extending the recent work of Kalai, Rothblum, and Rothblum (CRYPTO ’17). In particular, these families are sufficient to instantiate the Fiat-Shamir heuristic in the plain model for a natural class of interactive proofs. An interesting consequence of our results is a potential new avenue for bypassing black-box separations. In particular, proving (with necessarily non-black-box techniques) that parallel repetition amplifies the hardness of specific one-way functions — for example, all one-way permutations — suffices to directly bypass Simon’s impossibility result.