AI models accelerate software and cryptographic security audits

Experts have discussed how AI models can be leveraged to improve software and cryptographic security, noting that companies are already spending significant resources scanning codebases with the latest models. Dan Boneh highlighted that AI can accelerate the validation of new cryptographic assumptions; instead of waiting a decade for human cryptographers to analyze a new scheme, developers can run AI models for a month to test for vulnerabilities. If the models fail to break the system, it provides strong evidence of security.

The discussion noted that AI capabilities are becoming significantly cheaper, with estimates suggesting intellectual capability per dollar is 99% cheaper than a year ago. This cost reduction, combined with the finite nature of software bugs, suggests that AI could help asymptotically approach zero bugs in software. The experts posited that this defensive application of AI is more likely to improve overall security than to create new offensive threats, as vulnerabilities can be designed away or mitigated using these increasingly affordable tools.

The panelists argued that AI is a net positive for security, enabling faster and more thorough auditing of software and cryptographic systems than traditional human methods. This shift allows for more rigorous testing of complex systems, potentially reducing the window of vulnerability for critical infrastructure and applications.