Anthropic’s AI Model Uncovers Thousands of Security Flaws
Anthropic’s most sophisticated artificial intelligence model, known as Claude Mythos Preview, has detected thousands of cybersecurity vulnerabilities affecting all major operating systems and web browsers. Rather than releasing the model publicly, Anthropic has chosen to distribute it discreetly to organizations responsible for maintaining the security and stability of the internet.
Introducing Project Glasswing
This initiative, called Project Glasswing, involves collaboration with prominent launch partners including Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks. Access has also been extended to over 40 additional organizations that develop or maintain critical software infrastructure.
Anthropic has committed up to $100 million in usage credits for Mythos Preview and allocated $4 million in donations to open-source security groups to support this effort.
A Model Surpassing Traditional Benchmarks
Although Mythos Preview was not explicitly trained for cybersecurity, its enhanced abilities in code analysis, reasoning, and autonomy have enabled it to excel at identifying and exploiting software vulnerabilities. The model has reached a performance level that saturates existing security benchmarks, prompting Anthropic to focus on discovering zero-day vulnerabilities — previously unknown software flaws.
Among its discoveries are a 27-year-old bug in OpenBSD, recognized for its robust security, and a fully autonomous identification and exploitation of a 17-year-old remote code execution vulnerability (CVE-2026-4747) in FreeBSD. This exploit allows an unauthenticated attacker to gain full control over servers running Network File System (NFS).
Nicholas Carlini, a researcher at Anthropic, highlighted the model’s extraordinary capability to chain multiple vulnerabilities to create complex exploits, stating, “I’ve found more bugs in the last couple of weeks than I found in the rest of my life combined.”
Reasons for Withholding Public Release
Newton Cheng, Frontier Red Team Cyber Lead at Anthropic, explained, “We do not plan to make Claude Mythos Preview generally available due to its cybersecurity capabilities. Given the rapid pace of AI advancement, such capabilities could proliferate beyond responsible actors, potentially causing severe consequences for economies, public safety, and national security.”
This caution is grounded in reality. Anthropic previously revealed the first documented AI-driven cyberattack, where a Chinese state-sponsored group autonomously infiltrated approximately 30 global targets using AI agents to conduct tactical operations.
The company has also confidentially briefed senior U.S. government officials on Mythos Preview’s capabilities. The intelligence community is actively evaluating how this technology could impact both offensive and defensive cyber operations.
Supporting Open-Source Security
Project Glasswing also addresses the chronic security challenges faced by open-source software. Jim Zemlin, CEO of the Linux Foundation, noted that security expertise has traditionally been limited to organizations with large, dedicated teams, leaving open-source maintainers to manage security largely on their own.
To support these critical maintainers, Anthropic has donated $2.5 million to open-source initiatives Alpha-Omega and OpenSSF via the Linux Foundation, as well as $1.5 million to the Apache Software Foundation. These funds facilitate access to AI-powered cybersecurity vulnerability scanning at an unprecedented scale.
Future Plans and Industry Impact
Anthropic aims to eventually deploy Mythos-class models broadly but only after implementing rigorous safeguards. The company plans to test these protections with the upcoming Claude Opus model, which carries lower risk than Mythos Preview.
Meanwhile, competition in the AI cybersecurity space is intensifying. OpenAI recently released GPT-5.3-Codex, its first model classified as high-capability for cybersecurity under its Preparedness Framework. Anthropic’s cautious approach with Project Glasswing signals a growing industry trend favoring controlled deployment of advanced AI models rather than open public release.
Whether such standards will persist as AI capabilities continue to evolve remains uncertain.
Fonte: ver artigo original
Related Articles
Thousands of Rare Concert Recordings Uploaded to Internet Archive by Longtime Music Enthusiast
Meta Expands Renewable Energy Portfolio with 650 MW of New Solar Power to Support AI Operations
ChatGPT and Gemini Voice Bots Often Repeat Falsehoods, While Alexa Stays Silent
Google Tests Remy AI Agent for Gemini, Emphasizing User Control and Task Automation
