Anthropic Restricts Access to Advanced AI Model for Cybersecurity Collaboration

By Andrew Chapados

*Published: 2026-04-09*

Anthropic has announced that its new artificial intelligence model, Mythos — a variant of its Claude AI system — possesses advanced capabilities in identifying cybersecurity vulnerabilities. The company plans to limit initial access to 40 select organizations as part of Project Glasswing, an initiative aimed at strengthening defenses against AI-powered cyber threats.

Project Glasswing seeks to use Mythos to detect and address vulnerabilities in critical software infrastructure before the model's broader release. Anthropic describes the effort as a potential "industry change point" for cybersecurity practices. Logan Graham, head of Anthropic's vulnerability testing team, stated that the model can identify issues "well understood and findable by security researchers," while also uncovering others missed by decades of human experts and automated tools. In some instances, Mythos has generated exploits for these vulnerabilities, according to Graham.

Anthropic operates Claude, a proprietary AI chatbot used in various applications, including "agentic" tasks such as software development. The company will provide up to $100 million in API credits to participants, equivalent to the cost of extensive model usage, and has pledged $4 million in donations to open-source security organizations. Selected participants include major technology firms such as Amazon, Apple, Google, and Microsoft; infrastructure and cybersecurity companies like Broadcom, Cisco, CrowdStrike, Nvidia, and Palo Alto Networks; financial institution JPMorgan Chase; and the Linux Foundation.

The restricted access is intended to ensure that only organizations with demonstrated responsibility and capability use Mythos to proactively fix vulnerabilities, preparing critical systems for potential future attacks enabled by similar AI tools. Anthropic emphasizes that the project focuses on defensive applications to enhance overall cybersecurity resilience.

This approach echoes past caution by AI developers. In 2019, OpenAI withheld full public release of its GPT-2 model, citing risks of misuse for generating propaganda or misleading content, as reported by Wired. Subsequent models have since surpassed GPT-2's capabilities and been made publicly available.

Claude has gained prominence in the tech sector for tasks like building software and applications. However, the system has faced scrutiny over reported incidents. In recent months, Anthropic acknowledged instances where files containing proprietary code were inadvertently uploaded to public repositories, exposing portions of Claude's source code, according to reports from journalist Aaron Holmes and others. A blog draft also reportedly revealed internal code before being removed.

Additionally, Anthropic has pursued legal action against the U.S. federal government, challenging a designation of the company as a potential supply chain risk. The company has also drawn attention for its AI safety framework, including a "Constitutional AI" approach overseen by Jan Leike, who has affiliations with Effective Altruism organizations focused on AI risk mitigation.

Anthropic's decisions come amid ongoing debates about AI deployment and safety. Elon Musk has publicly criticized Claude for responses to certain prompts, as noted in prior coverage. The company continues to position itself as prioritizing responsible AI development, with Mythos representing a step toward industry-wide cybersecurity improvements through controlled collaboration.

Project Glasswing participants are expected to collaborate on patching vulnerabilities in open-source and critical infrastructure software, contributing findings back to the community. Anthropic has not specified a timeline for Mythos's public availability but indicated that the initiative will inform future safeguards.

As AI models advance, companies like Anthropic are balancing innovation with risk management. Graham noted that Mythos's abilities highlight the dual-use potential of such technology: powerful for defense but requiring safeguards against offensive applications.

(Word count: 652)