The AI startup Anthropic has unveiled Project Glasswing, a cybersecurity initiative that unites prominent technology companies and critical infrastructure entities to evaluate a new, unreleased AI model aimed at detecting and addressing software vulnerabilities on a large scale.
This initiative centers around an advanced AI system equipped with robust coding and analytical capabilities. Anthropic asserts that this model can scrutinize significant risks within operating systems, web browsers, and commonly used software. Initial testing has allegedly revealed numerous serious vulnerabilities, some of which had remained unrecognized for years.
In contrast to conventional AI releases, the model—known among collaborators as Claude Mythos Preview—will not be made available to the general public due to concerns about potential misuse. Access is restricted to a select group of partners within the technology and security sectors.
Microsoft, one of the participating companies, highlighted that this initiative signifies a transformation in the approach to cybersecurity on a larger scale. Igor Tsyganskiy, Global CISO and EVP of Security and Microsoft Research at Microsoft, stated, “As we enter a phase where cybersecurity transcends human limitations, the opportunity to leverage AI responsibly for enhanced security and risk reduction is unparalleled.”
He added, “Participating in Project Glasswing enables us to identify and address risks at an early stage. When evaluated against CTI-REALM, our open-source security benchmark, Claude Mythos Preview demonstrated significant advancements over prior models.”
Amazon Web Services (AWS) reported that it has already started integrating the model into its internal systems. Amy Herzog, Vice President and CISO at AWS, remarked, “At AWS, we prioritize establishing defenses before threats materialize. AI is crucial to our capacity to defend on a large scale.”
She further noted, “We’ve been utilizing Claude Mythos Preview in our security operations, where it is already enhancing our code security.”
Meanwhile, Google underscored the necessity of collaborative industry efforts to address emerging cybersecurity threats. Heather Adkins, VP of Security Engineering at Google, expressed satisfaction with the formation of this cross-industry initiative, stating, “It has always been vital for the industry to collaborate on new security challenges.”
She added, “We have consistently believed that AI presents both new challenges and new opportunities in the realm of cyber defense.”
According to Anthropic, the primary aim of Project Glasswing is to bolster defensive cybersecurity measures in light of increasingly sophisticated AI-driven threats. The model can pinpoint vulnerabilities and, in some instances, aid in understanding how these weaknesses might be exploited, thus highlighting its defensive advantages as well as inherent risks.
The initiative also seeks to enhance security within open-source software, which is foundational to much of today’s digital infrastructure but often lacks adequate resources. By offering access to cutting-edge AI tools, Anthropic intends to assist maintainers in proactively identifying and rectifying vulnerabilities.
The company has pledged up to $100 million in usage credits and $4 million towards open-source security initiatives, with plans to broaden participation to additional organizations responsible for essential systems.
This launch occurs amid escalating concerns regarding the surge in AI-driven cyberattacks, with experts cautioning that increasingly sophisticated models could hasten both the discovery and exploitation of vulnerabilities. Anthropic contends that Project Glasswing represents an initial effort to ensure that such capabilities are utilized responsibly and primarily for defensive objectives.
