What Is Project Glasswing?

Project Glasswing is an Anthropic research initiative focused on developing AI models with substantially higher intelligence than the current Opus line — comparable to or exceeding human expert-level capability in specialized domains. The project name was first mentioned in Anthropic’s research updates, and the Claude Opus 4.8 announcement provided the most concrete public update on its status.

Anthropic describes Glasswing as encompassing both the development of more capable models and the safety infrastructure needed to deploy them responsibly. The project’s dual focus on capability and safety is consistent with Anthropic’s broader mission of building AI systems that are both powerful and aligned with human interests.

Claude Mythos Preview

Claude Mythos Preview is the first model to emerge from Project Glasswing. It is currently available in a limited preview to a small number of organizations, specifically for cybersecurity work.

The term “Mythos” suggests a new model class above Opus in Anthropic’s naming hierarchy. While Opus is Anthropic’s most capable publicly available model, Mythos appears to represent a higher tier of intelligence — one that requires additional safety measures before broad deployment.

Anthropic’s alignment assessment for Opus 4.8 revealed that its alignment quality is similar to Claude Mythos Preview, which the company describes as its “best-aligned model.” This is notable because it suggests that Opus 4.8 benefits from safety techniques originally developed for the Mythos line.

Why Cybersecurity?

The cybersecurity focus for Mythos Preview is strategic. Cybersecurity is a domain where:

• Expert-level AI could have immediate impact — threat detection, vulnerability analysis, and incident response are all areas where faster, more intelligent analysis can make a tangible difference
• Safety is paramount — a powerful AI model in cybersecurity must be carefully controlled to prevent misuse
• Real-world testing is feasible — cybersecurity organizations can evaluate the model’s effectiveness in controlled environments

By starting with cybersecurity, Anthropic can validate Mythos-class capabilities in a high-stakes domain while the safety infrastructure is refined for broader use cases.

The Cyber Safeguard Challenge

Anthropic has explicitly stated that “models of this capability level require stronger cyber safeguards before they can be generally released.” This is a critical aspect of Project Glasswing’s mission.

The concern is straightforward: more capable AI models are also more capable dual-use tools. A model with expert-level cybersecurity knowledge could be used to defend systems — or to attack them. Anthropic is investing in safeguards that prevent or substantially mitigate the potential for misuse while preserving the model’s beneficial capabilities.

These safeguards likely include:

• Refusal training to reject malicious cybersecurity requests
• Usage monitoring to detect and prevent abuse patterns
• Access controls for organizations using the model
• Red-teaming to identify and address vulnerabilities before release

Anthropic’s progress on these safeguards will determine when Mythos-class models can be made more broadly available.

Mythos vs Opus: How They Compare

Based on available information, here is how Mythos Preview compares to the Opus line:

The alignment comparison is particularly interesting. The fact that Opus 4.8 has alignment quality “similar to our best-aligned model, Claude Mythos Preview” suggests that techniques developed for the more capable model have been applied to improve Opus 4.8 as well.

Release Timeline & Availability

Anthropic stated in the Opus 4.8 announcement that they are “making swift progress on developing these [cyber] safeguards and expect to be able to bring Mythos-class models to all our customers in the coming weeks.”

This timeline suggests:

• Current state: Limited preview with select organizations for cybersecurity
• Coming weeks: Broader availability of Mythos-class models, potentially including general-purpose capabilities
• Beyond: Continued development of even more capable models through Project Glasswing

It’s important to note that “coming weeks” is an optimistic timeline from Anthropic’s announcement and may shift based on safety testing results.

Industry Implications

Project Glasswing and Claude Mythos Preview have several implications for the AI and cybersecurity industries:

• AI capability ceiling is rising — Mythos represents a new tier of AI intelligence above the current Opus line, suggesting further rapid advancement
• Safety is a release gating factor — Even for the most capable AI labs, safety infrastructure development is now the bottleneck for deploying advanced models, not capability development alone
• Cybersecurity is a proving ground — Success in the cybersecurity domain could pave the way for Mythos-class models in other expert domains like legal, medical, and financial analysis
• Competitive pressure — As OpenAI, Google, and others race toward more capable models, Anthropic’s focus on safety-aligned capability could become a differentiator or a constraint depending on how quickly safeguards can be developed

Frequently Asked Questions