Anthropic, the AI safety and research company, announced the launch of a new frontier AI model, Claude Mythos Preview. However, the access is controlled under ‘Project Glasswing’ to secure critical software in the AI era.
Security vulnerabilities led to a controlled Claude Mythos Preview
A recently discovered wave of “high-security vulnerabilities” across operating systems, web browsers, and other core software has led Anthropic to bring on a swift decision to restrict access to the Claude Mythos Preview model.
The model will now only be available to a “selected group of companies” as a precautionary measure against exploitation. The company stated that the main goal is to deploy the Mythos-class model safely while ensuring methods to block their most dangerous outputs.
Mythos Preview is capable of identifying previously unknown security flaws called ‘zero-days’ and then exploiting them in every major operating system and web browser under the direction of a user.
According to media reports, hackers have been using AI for cyberattacks. There has been a 72% year-over-year increase in AI-powered cyberattacks, with 87% of global organizations reporting AI-enabled cyberattacks in 2025.
Anthropic has been concerned about these concerns especially on the consequences of similar AI capabilities being used by bad actors. The company developed and launched a solution called Project Glasswing to get ahead of threats by fixing critical vulnerabilities before bad actors exploit them.
Project Glasswing to secure AI
The initiative, according to the official report, will bring together 40 major giants in the industry including Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks to secure the world’s ‘most critical software’.
“We formed Project Glasswing because of capabilities we’ve observed in a new frontier model trained by Anthropic that we believe could reshape cybersecurity,” Anthropic said in their report.
The company viewed this initiative only as a beginning, claiming that more organizations need to work together to solve the concerns of cybersecurity, including AI developers, other software companies, security researchers, open-source maintainers, and regulators.
The report also discussed how the role of bugs has always been a part of the software equation. However, their impact could vary based on how serious of a flaw they could lead to. Some could allow cyberattackers to hijack systems, disrupt operations, or steal data.
The company used Claude Mythos Preview to identify these flaws, including a 27-year-old vulnerability in OpenBSD, a security-hardened operating system used in running firewalls, and a 16-year-old vulnerability in FFmpeg, a multimedia framework for decoding, encoding, and more.
The model even identified and chained together several vulnerabilities in Linux Kernel, the core of any Linux-based operating system used globally, that could give one user the access to control the whole system. Chaining vulnerabilities is much more powerful than exploiting just one and is significant to the nature of Linux’s operations.
The company has been researching various ways to train the model on emotion concepts and their functions in Large Language Models (LLM), and reported how they observed emotion vectors in autonomous AI models. Anthropic identified that under various simulated scenarios, 16 major AI models had misaligned behaviors.
