Anthropic Brings Apple, Google and 45 Other Organizations Together to Defend AI Against Cyberattacks

Anthropic has announced the launch of Project Glasswing, a large-scale industry coalition that includes Apple, Google and more than 45 other companies and organizations. The goal is to work out together what modern AI systems can do in cybersecurity, and to develop practical approaches to both offensive and defensive uses of frontier artificial intelligence.

The coalition comes amid growing concern in the industry: many experts believe the development of offensive AI capabilities is clearly outpacing the development of defensive tools.

The coalition's key tool will be Claude Mythos Preview, a new specialized model from Anthropic built specifically for information security tasks. Unlike the commercial versions of Claude, Mythos Preview is designed for in-depth work with vulnerabilities: automatically finding weaknesses in software systems, analyzing malicious code, and assessing how far an advanced AI system can get in a targeted attack on digital infrastructure. At the same time, coalition members will test how the same capabilities work in a defensive mode: detecting anomalies, responding to incidents and proactively protecting systems.

The lineup of participants itself shows how serious the moment is. Apple and Google, direct competitors of Anthropic in the large language model market, have agreed to work side by side on one project. That rarely happens in a tech industry where the fight for market share is zero-sum. Cybersecurity has turned out to be one of the few areas where competition between companies temporarily gives way to a common threat: vulnerabilities in AI systems do not respect corporate boundaries, and the consequences of potential attacks affect the entire digital ecosystem.

The project rests on an alarming but already obvious observation: modern AI systems easily automate tasks that five years ago required a skilled specialist and many weeks of work, such as systematic vulnerability discovery, writing working exploits and simulating targeted attacks. Some researchers have already documented cases where AI agents successfully solved basic hacking tasks and found zero-day vulnerabilities in real software. As models grow more powerful and more accessible, this threshold will keep falling. Project Glasswing is an attempt to measure the real scale of the threat in advance, before attackers start using these capabilities at full strength.

The initiative also has a clear strategic dimension for Anthropic. The company has consistently built its reputation on the principle of responsible AI development, regularly naming safety as its top priority. Project Glasswing turns declarations into concrete action: instead of studying risks in a closed lab, Anthropic is inviting the whole market, competitors included, to work together. This strengthens the company's position as a venue for shaping industry standards and norms in AI security, where Anthropic clearly aims for leadership.

Technical details about Claude Mythos Preview remain undisclosed. Anthropic has not revealed the architecture, the parameter count or any comparative benchmarks. All that is known is that it is a separate specialized version outside the commercial Claude lineup, and that coalition members will get controlled access to it for their own research. The results of the joint work are meant to feed into practical recommendations and, possibly, cross-industry standards for using AI in cybersecurity.

Project Glasswing marks an important shift in how the technology industry approaches AI and security. The question of whether AI systems can become an attack tool is already settled: in the hands of sufficiently motivated actors, they can and already have. The open question is now a different one: will the industry manage to build reliable defense mechanisms before these threats grow into a systemic crisis for global digital infrastructure?