Anthropic’s Claude Mythos Breakthrough: ‘Project Glasswing’ Launched to Prevent AI Cyber-Crisis

The same capabilities that make AI more powerful may also be sharpening the tools of cyberattacks, and that worry is now impossible to ignore.

After testing its latest frontier model, ‘Claude Mythos Preview’, Anthropic has released a note of caution. It stated the model doesn’t just write code better, it can also find and exploit its weaknesses with alarming efficiency.

In a newly released document, Anthropic warned such capabilities could make cyberattacks faster, cheaper and significantly more damaging, pushing them closer to (or even beyond) human-level sophistication.

The implications stretch far beyond the tech world, touching everything from economic stability to national security.

To curb this, Anthropic has launched ‘Project Glasswing’, a cross-industry initiative designed to reinforce the foundations of software security before vulnerabilities can be weaponised at scale.

Stated Problem & Proposed Solution

Anthropic’s problem statement is that modern AI models are now capable of analysing code deeply enough to uncover flaws that might otherwise go unnoticed. Without strong safeguards, this could dramatically lower the barrier to launching high-impact cyberattacks.

Project Glasswing aims to counter that threat by bringing together a coalition of tech and enterprise leaders. It includes Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, to collectively harden critical software infrastructure.

According to Anthropic, the initiative is a direct response to the cybersecurity risks revealed by Claude Mythos Preview itself.

What is Claude Mythos Preview?

Claude Mythos Preview is Anthropic’s new general-purpose language model that demonstrates strong performance across tasks, with particularly notable capabilities in computer security tasks.

According to the start-up, the model has already identified thousands of zero-day vulnerabilities. It supposedly uncovered critical flaws across major operating systems and browsers, including a 27-year-old bug in OpenBSD, a 16-year-old issue in FFmpeg, and a chain of Linux kernel vulnerabilities that could enable full system control.

Mythos Capabilities

Anthropic presented benchmark data indicating that Mythos Preview outperforms its previous model, Claude Opus 4.6, in coding and reasoning tasks.

In its announcement blog, the start-up said the model achieved higher scores on benchmarks such as CyberGym vulnerability reproduction, SWE-bench Pro, Terminal-Bench 2.0, SWE-bench Multilingual, SWE-bench Verified, GPQA Diamond, Humanity’s Last Exam, BrowseComp and OSWorld-Verified.

It also noted certain caveats, including memorisation filters and the use of an internal implementation for one multimodal benchmark.

Anthropic said it does not plan to release Claude Mythos Preview for general use. Instead, it intends to use the model to develop cybersecurity safeguards that can enable the safe deployment of similar “Mythos-class” systems at scale. The company also indicated that it plans to introduce new safeguards with an upcoming Claude Opus model.

Project Glasswing

Project Glasswing is a new initiative that brings together several big tech companies in an effort to secure the world’s most critical software.

It will focus on practical security applications. Anthropic said partners will use the model for tasks such as local vulnerability detection, black-box testing of binaries, endpoint security and penetration testing of foundational systems.

The broader goal is to help secure a significant portion of the global cyberattack surface.

Anthropic is supporting the initiative with funding, credits, and expanded access. The company said it will provide up to $100 million in model usage credits for Project Glasswing and related efforts, along with donations to open-source security groups.

Following the research preview phase, the model is expected to be accessible through platforms such as the Claude API, Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry at standard token-based pricing.

Anthropic said it aims for the project to generate shared industry guidance. The company plans to publish a public report within 90 days outlining key findings and disclosures, and to collaborate with security groups on recommendations covering vulnerability disclosure, patching, supply chain security, secure-by-design development, standards for regulated industries, and automation.