AnthropicのProject Glasswingはモデル悪用防止に不十分かもしれない

3 Min ReadAnthropic's introduction of Project Glasswing is an example of ways it's continuing to uphold its reputation as a responsible generative AI vendor, but also highlights the growing cyberscurity danger of models becoming better at human tasks such as code generation.The Claude maker launched the project on April 7 as an initiative that brings together major companies, including AWS, Apple, Nvidia, JPMorgan Chase and Palo Alto Networks. The project aims to provide security for what the vendor called "the world's most critical software."Project Glasswing comes nearly two weeks after a data leak released limited details about Mythos to the public. The model itself highlights an important trend in the AI market: cybersecurity is now a key application. However, this trend has been a long time coming because of the rapid growth of AI models, which has also led to increased security risks, such as deepfakes, that have affected elections and even organizations.Related:The Real AI Shift Isn’t New Models. It’s Control.Anthropic's Concerns and Responsible AIAnthropic said it formed the project after observing Claude Mythos, its unreleased model that is now in preview. Claude Mythos is proof that AI models have achieved a coding capability that surpasses humans' ability to find and exploit software vulnerabilities, Anthropic said. The vendor said Claude Mythos has already found thousands of vulnerabilities in every major OS and web browser, warning that if the capabilities spread and fall into the hands of bad actors, the consequences could be severe for economies, public safety and national security.Therefore, the partners in Project Glasswing will use Mythos Preview as a defense mechanism. The AI lab said it plans to share what the project learns and will extend access to 40 other organizations that build software. Anthropic is also talking with U.S. government officials about Claude Mythos Preview and ways it can contribute to offensive and defensive cyber capabilities.Project Glasswing appears to be an attempt by Anthropic to demonstrate that, despite downgrading its Responsible Scaling Policy earlier this year, it remains committed to responsible AI."In this context, responsible AI should be understood as minimizing aggregate societal risk from a dual-use capability," said RPA2AI Research CEO Kashyap Kompella, referring to the idea that a technology can be used for both offensive and defensive hacking. "On that definition, restricted release is more responsible than public release."The Problem PersistsGlasswing provides defenders first access so they can "harden foundational systems before adversaries obtain comparable tools," Kompella added.Related:OpenAI GPT-5.4-Cyber is More Open Than Claude Mythos"Such a coalition could help establish new norms for model release, vulnerability triage, patch-cycle compression, and security benchmarking before cyber-capable models become widespread," he said. "This is a mitigation strategy, not a resolution."The problem of advanced AI models like Mythos potentially causing havoc if they fall into the wrong hands persists because new, better models are released almost daily. The industry is moving so fast, and code generation and autonomy continue to improve with new releases, Kompella added.However, the good news for cybersecurity firms is that, while automated vulnerability discovery may lead to more vulnerabilities, they still have opportunities in their validation, prioritization, patch orchestration and compliance translation processes.About the AuthorNews Writer, AI BusinessEsther Shittu brings four years of expertise covering artificial intelligence technologies and industry trends. As co-host of the "Targeting AI" podcast, she talks to thought leaders and practitioners exploring critical AI developments. Previous to AI Business, she wrote for several publications including the New York Daily News, Bklyner and the Brooklyn Daily Eagle. When she's not diving deep into the world of AI, she spends her time on passion projects and raising her three daughters.