Back to list
Industry NewsAnthropicCybersecurityArtificial Intelligence

Project Glasswing: Anthropic Partners with Tech Giants to Secure Critical Software Against AI-Driven Cyber Threats

Anthropic has announced Project Glasswing, a major cybersecurity initiative involving industry leaders such as Amazon Web Services, Apple, Google, Microsoft, and NVIDIA. The project is a response to the capabilities of Claude Mythos Preview, a new unreleased frontier model that has demonstrated the ability to surpass most humans in finding and exploiting software vulnerabilities. Mythos Preview has already identified thousands of high-severity vulnerabilities across major operating systems and web browsers. To combat the potential risks of AI-driven exploits, Anthropic is committing $100 million in usage credits and $4 million in donations to open-source security organizations. The initiative aims to leverage these advanced AI capabilities for defensive purposes, securing both first-party and open-source infrastructure before such tools proliferate to malicious actors.

Hacker News
Share

Key Takeaways

  • Unprecedented AI Capabilities: Anthropic’s unreleased model, Claude Mythos Preview, can find and exploit software vulnerabilities more effectively than almost all human experts.
  • Broad Industry Coalition: Major tech players including AWS, Apple, Google, Microsoft, NVIDIA, and Cisco have joined Project Glasswing to bolster defensive security.
  • Critical Vulnerabilities Found: The Mythos Preview model has already discovered thousands of high-severity bugs in every major operating system and web browser.
  • Significant Financial Commitment: Anthropic is providing $100 million in model credits for defensive work and $4 million in direct donations to open-source security groups.

In-Depth Analysis

The Emergence of Claude Mythos Preview

The catalyst for Project Glasswing is the development of Claude Mythos Preview, a general-purpose frontier model that represents a significant leap in coding proficiency. According to Anthropic, this model has reached a threshold where it can identify and exploit software vulnerabilities with a level of skill that rivals or exceeds the most elite human cybersecurity professionals. The model's performance has already been proven in real-world scenarios, having identified thousands of high-severity vulnerabilities within the core infrastructure of the modern digital world, including every major web browser and operating system.

Defensive Mobilization and Industry Collaboration

Recognizing that these AI capabilities will eventually proliferate, Project Glasswing serves as an urgent defensive measure to secure critical software before malicious actors can utilize similar technology. The initiative brings together a powerhouse coalition of launch partners, including Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. These organizations will utilize Mythos Preview to scan and secure their systems. Furthermore, Anthropic has extended access to over 40 additional organizations responsible for maintaining critical software infrastructure, ensuring that both proprietary and open-source systems are reinforced against potential AI-driven attacks.

Resource Allocation for Open-Source Security

Anthropic is backing this initiative with substantial financial resources to ensure the security benefits reach beyond large corporations. The company is committing up to $100 million in usage credits for Mythos Preview to facilitate defensive security research and implementation. Additionally, $4 million in direct donations will be distributed to open-source security organizations. This move highlights the importance of protecting the open-source ecosystem, which forms the backbone of global digital infrastructure but often lacks the resources of major tech enterprises.

Industry Impact

Project Glasswing marks a pivotal moment in the AI industry where the focus shifts from general productivity to high-stakes cybersecurity defense. By proactively identifying vulnerabilities in major operating systems and browsers, the initiative could significantly reduce the attack surface available to cybercriminals. However, it also signals a new era of "AI vs. AI" security, where the speed of software patching must keep pace with the speed of AI-driven discovery. The collaboration between traditional competitors like Google, Microsoft, and Apple suggests that the threat posed by frontier AI models is viewed as a systemic risk that requires a unified industry response.

Frequently Asked Questions

Question: What is Project Glasswing?

Project Glasswing is a collaborative initiative led by Anthropic and several major technology companies to use advanced AI models for defensive cybersecurity, specifically to find and fix vulnerabilities in critical software.

Question: What makes the Claude Mythos Preview model unique?

Claude Mythos Preview is an unreleased frontier model that has demonstrated the ability to find and exploit software vulnerabilities at a level that surpasses almost all human experts, having already found thousands of bugs in major operating systems.

Question: Who are the partners involved in this project?

Launch partners include AWS, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, along with over 40 other organizations.

Read original article

Related News

Protecting Engineering Expertise: Why AI Efficiency Could Threaten the Next Generation of Specialists
Industry News

Protecting Engineering Expertise: Why AI Efficiency Could Threaten the Next Generation of Specialists

In a thought-provoking analysis, Richard Mitchell, systems engineer and CEO of AuraSpark Technologies, warns that the rapid pursuit of AI efficiency may come at a significant cost: the erosion of human expertise. Drawing critical parallels from the aviation and nuclear power industries, Mitchell highlights the dangers of over-reliance on automation. As AI takes over complex engineering tasks, there is a growing concern that the next generation of experts will lack the foundational skills and hands-on experience necessary to manage systems when technology fails. The article emphasizes that preserving human skill sets is not just a matter of professional development, but a safety-critical necessity in high-stakes environments. This shift requires a strategic balance between leveraging AI for productivity and ensuring that human oversight remains robust and informed by deep technical knowledge.

Benchmarking AI Coding Agents: A Deep Dive into Tool Selection Across 17,000 Experimental Runs
Industry News

Benchmarking AI Coding Agents: A Deep Dive into Tool Selection Across 17,000 Experimental Runs

A comprehensive study has analyzed how prominent AI coding agents, including Claude, Codex, and Cursor, select third-party tools and services during software development tasks. By analyzing thousands of public GitHub repositories, researchers established a balanced panel of 75 repositories across 10 different programming languages, utilizing real-world statistics to ensure the data was not biased toward open-source startups. The experiment employed four distinct developer personas—Vibe-coder, Junior engineer, Senior engineer, and Enterprise engineer—to test how varying levels of professional requirement and constraint affect AI decision-making. With 1,163 prompt variations and thousands of runs conducted in ephemeral sandboxes, the study provides a rigorous framework for understanding the logic and preferences of AI agents when tasked with implementing features like email services or invoice generation in complex codebases.

Cerebras Inference Platform Achieves Record Speeds with Qwen 3.8 27B and OpenAI GPT OSS 120B
Industry News

Cerebras Inference Platform Achieves Record Speeds with Qwen 3.8 27B and OpenAI GPT OSS 120B

Cerebras Systems has announced a significant performance update to its inference platform, featuring the Qwen 3.8 27B and OpenAI GPT OSS 120B models. According to the latest documentation, the Qwen 3.8 27B model now operates at approximately 1500 tokens per second, while the GPT OSS 120B model reaches an impressive 3000 tokens per second. These models are available through various access tiers, including free trials and pay-as-you-go options, with context windows extending up to 131k. A key highlight of this release is Cerebras' commitment to model quality; all models served via public endpoints are unpruned versions. The platform utilizes selective weight-only quantization for storage to maintain high precision during operations, ensuring that quality-sensitive layers remain at full precision through on-the-fly dequantization.