Mint Explainer | Can AI find bugs humans can’t? Inside Anthropic’s Project Glasswing

Mint explains how.Anthropic cofounder and chief executive Dario Amodei said the company created a cybersecurity coalition after observing how capable its new ‘frontier’ artificial intelligence model, Claude Mythos, is. Anthropic tested the latter to discover thousands of critical software flaws, technically known as ‘zero-day vulnerabilities’, within an unspecified period of time.A ‘zero-day’ vulnerability means a glitch in a software’s code that no one has discovered to date.

In the cybersecurity world, such a glitch is the hardest to detect and often requires the most skilled cybersecurity engineers.Anthropic noted that cyber attackers worldwide thrive on such flaws and can bring down large enterprises and even governments. Two of the most notable such instances include June 2010’s ‘Stuxnet’, which targeted Iran’s nuclear research facilities, and 2017’s WannaCry and NotPetya cyberattacks, which exploited a zero-day exploit tool used by the US government’s National Security Agency (NSA).No.

Project Glasswing, at its core, is an AI model that can find gaps in the most commonly used software, and write the code for making software patches for the same.Think of this as a layer of research that finds undiscovered security flaws and vulnerabilities in software such as Apple’s, Microsoft’s and Google’s iOS, Windows and Android operating systems. So far, this research has largely been conducted by human-led cybersecurity discovery teams at companies such as Cisco, CrowdStrike, Fortinet, and Palo Alto Networks.Cybersecurity applications use this foundational research data to update their repository of software patches and knowledge of vulnerabilities in software systems, and thus enable user-facing services that build on tools