Anthropic has launched Project Glasswing, a coalition of more than 45 organizations — including direct competitors Apple and Google — to evaluate its new Claude Mythos Preview model against cutting-edge cybersecurity threats.

The initiative represents one of the broadest cross-industry collaborations in AI safety to date. Rather than competing in isolation, the participating companies will pool their efforts to stress-test how advanced AI models perform in cybersecurity contexts — both as potential defensive tools and as systems that could be exploited to carry out attacks.

Rivals at the Same Table

The decision to include competitors in a shared testing framework is notable. Anthropic, Apple, and Google are all active players in the AI model market, yet Project Glasswing places them in a cooperative structure under Anthropic's coordination. The breadth of the coalition — spanning more than 45 organizations according to Anthropic — suggests the company views cybersecurity risk as a threat that no single lab or corporation can adequately address alone.

This represents one of the broadest cross-industry AI safety collaborations to date, uniting direct competitors around a shared concern that transcends market rivalry.

The new Claude Mythos Preview model sits at the center of the effort. While Anthropic has not disclosed the full technical specifications of Mythos Preview, the model appears designed with cybersecurity evaluation in mind — capable of reasoning about vulnerabilities, attack vectors, and defensive strategies at a level of sophistication that warrants coordinated external scrutiny before any wider deployment.

Why Cybersecurity, Why Now

AI-assisted cyberattacks are no longer a theoretical concern. Security researchers have already demonstrated that large language models can assist in writing malicious code, identifying software vulnerabilities, and crafting convincing phishing content. A 2023 study by researchers at ETH Zurich found that GPT-4 could autonomously exploit one-day vulnerabilities in real systems when given access to relevant documentation — a finding that accelerated industry attention to the offensive potential of frontier models.

For Anthropic, Project Glasswing appears to be a direct response to that escalating risk landscape. By recruiting organizations with deep cybersecurity expertise and operational infrastructure, the company gains access to testing environments and threat intelligence that its own team could not replicate internally. Participants will use Claude Mythos Preview to probe where advanced AI capabilities intersect most dangerously with known and emerging attack surfaces.

What the Coalition Structure Signals

The multi-stakeholder format of Project Glasswing echoes other industry-wide safety initiatives, such as the Frontier Model Forum launched in 2023 by Anthropic, Google, Microsoft, and OpenAI. That body similarly brought together competing labs to share safety research and establish shared norms. Project Glasswing follows the same logic but applies it to a specific and urgent technical domain.

The involvement of Apple is particularly significant. Apple has historically been cautious about external AI partnerships and is not typically associated with frontier model development in the way Google or Anthropic are. Its participation suggests the coalition's remit may extend into device-level and operating system security, where AI-enabled attacks could target consumer hardware at scale.

For individuals and organizations that depend on digital infrastructure — which is to say virtually everyone — the stakes are concrete. A successful AI-assisted cyberattack on critical systems, financial networks, or healthcare infrastructure could cause harm far beyond what traditional hacking tools enable. The speed and adaptability of AI models could allow attackers to probe defenses at a pace human security teams cannot match.

What Comes After Testing

It is not yet clear what happens to the findings generated through Project Glasswing. Whether results will be published, shared privately among participants, or used solely to inform Anthropic's own model development remains undisclosed, according to the information available from Wired's reporting. The lack of a stated publication commitment is a meaningful gap — collaborative safety work has the greatest impact when findings are shared with the broader security community, not retained as proprietary intelligence.

Anthropics track record on transparency is mixed. The company publishes regular model cards and safety evaluations, but like its peers, it treats core model architecture and training data as confidential. How much of what Project Glasswing learns becomes public knowledge will determine how much the broader industry benefits.

The timeline for Project Glasswing's evaluation phase and any subsequent deployment decisions for Claude Mythos Preview has not been announced, according to Anthropic.

What This Means

For businesses and policymakers tracking AI risk, Project Glasswing is a concrete signal that leading technology companies now treat AI-enabled cyberattacks as an immediate operational threat — serious enough to set aside competitive interests and test their models together before those capabilities reach the open market.