Vulnerability signatures offer better precision and flexibility than exploit signatures when detecting network attacks. We show that it is possible to detect vulnerability signatures in high-performance network intrusion detection systems, by developing a matching architecture that is specialized to the task of vulnerability signatures. Our architecture is based upon: i) the use of high-speed pattern matchers, together with control logic, instead of recursive parsing, ii) the limited nature and careful management of implicit state, and iii) the ability to avoid parsing large fragments of the message not relevant to a vulnerability. We have built a prototype implementation of our architecture and vulnerability specification language, called VESPA, capable of detecting vulnerabilities in both text and binary protocols. We show that, compared to full protocol parsing, we can achieve 3x or better speedup, and thus detect vulnerabilities in most protocols at a speed of 1 Gbps or more. Our architecture is also well-adapted to being integrated with network processors or other special-purpose hardware. We show that for text protocols, pattern matching dominates our workload and great performance improvements can result from hardware acceleration.