Anthropic Expands Project Glasswing, Introduces Three-Tier Access

Integration with Cyber Verification Program enhances AI cybersecurity collaboration across 200 global institutions

Anthropic is expanding its cybersecurity collaboration 'Project Glasswing,' which supports the use of artificial intelligence (AI) models in cybersecurity.

Anthropic announced on the 6th (local time) that it will integrate Project Glasswing and the Cyber Verification Program (CVP), which have been operated separately until now.

Project Glasswing is a program that provides access to Claude Mythos, the top-tier AI model. CVP has supported the use of lower-tier models, 'Claude Opus' and 'Claude Sonnet,' in the cybersecurity field.

The new access authorization system is divided into three tiers: 'Defensive Access,' 'Red Team Access,' and 'Special Access.'

Defensive Access targets public institutions, companies, and universities engaged in defensive tasks such as security center operations and vulnerability analysis.

Under Red Team Access, users can perform simulated system intrusions and attack tests. However, actions that could cause large-scale disruptions are restricted.

Special Access, the top tier, is provided only to select verification agencies testing safety measures for critical infrastructure such as power grids, air traffic operations, communication networks, interbank transfer systems, and government administrative networks.

Existing Project Glasswing participants will automatically transition to the Special Access tier without additional procedures. New applicants will be verified jointly by Anthropic and the U.S. government.

Project Glasswing is a cybersecurity collaboration program launched by Anthropic in April to identify and address vulnerabilities in major software and critical infrastructure using AI.

It began with about 50 institutions, including Amazon Web Services (AWS), Apple, Google, Microsoft, NVIDIA, Cisco, CrowdStrike, and JPMorgan Chase, and expanded in June by adding approximately 150 institutions from over 15 countries, bringing the total to about 200 participants. Domestically, participants include the Korea Internet & Security Agency, KISA, Samsung Electronics, SK Hynix, and SK Telecom.

From April to July, over 129,000 software vulnerabilities were identified through Project Glasswing, with 33,000 cases classified as critical or high-risk.

However, there are concerns about the enhanced cybersecurity capabilities of AI models. Jamie Dimon, CEO of JPMorgan Chase, stated in an interview with Bloomberg TV on the same day, "Anthropic's Mythos model has increased global cybersecurity risks tenfold."