Anthropic PBC is expanding a program that enables organizations to test their cybersecurity defenses using its Claude Mythos Preview model.

The initiative, which is known as Project Glasswing, launched earlier this year with about 50 participants. The group included Microsoft Corp., Nvidia Corp., Google LLC and prominent tech firms. Anthropic announced today that it has opened access to 150 more organizations.

The company says the newly added participants include enterprises in the power, water, healthcare and communications sectors. The Financial Times reported that NATO and the European Union’s ENISA cybersecurity agency also received access.

Anthropic is limiting Claude Mythos Preview’s availability over concerns that it could be misused by hackers. However, some industry observers argue that the model’s phased rollout doesn’t fully mitigate cybersecurity risks. 

“Controlled rollout of frontier AI is the right instinct. But opacity is not a security strategy,” said Justin Beals, founder and chief executive of cybersecurity startup Strike Graph Inc. “As these tools become more capable, the organizations cleared to use them become high-value targets. Access without continuous compliance validation is just a slower version of the same risk. Whoever gets access, the standard should be verifiable transparency, not curated receipts.”

One of the reasons behind Anthropic’s cautious approach is that Claude Mythos Preview can not only spot software vulnerabilities but also find ways to exploit them. In some cases, it can chain together multiple vulnerabilities. Commercially available large language models struggle with that task.

Last month, Anthropic disclosed that Project Glasswing participants had found more than 23,000 vulnerabilities. Claude Mythos Preview rated more than a quarter of those flaws as high-severity or critical. Anthropic manually analyzed 1,752 of the flaws that receive those destinations and determined that 90.6% of them indeed qualify.

According to the company, Project Glasswing participants are using Claude Mythos Preview not only to find vulnerabilities but also to fix them. The model has proven adept at writing cybersecurity patches. In addition, some early adopters are using it to scan new code for weak points before rolling it out to production.

The lessons that Anthropic gleaned from Project Glasswing informed the development of a tool it describes as a threat model builder. According to the company, it helps LLMs prioritize the vulnerable parts of a code base when running cybersecurity scans. Last month, Anthropic started giving some Claude customers access to the threat model builder along with certain other technical resources developed as part of Project Glasswing.

Going forward, the company plans a significant increase in the number of participants in the cybersecurity program. It will bring in organizations from both the U.S. and abroad. Project Glasswing already includes participants from more than a dozen countries.

Separately, Anthropic will ramp up its efforts to use Claude Mythos Preview to fix vulnerabilities in open-source projects. The company plans to share best practices for sharing vulnerabilities with open-source project maintainers. In the longer term, Anthropic intends to develop industry standards and other technical resources designed to address the cybersecurity risks posed by advanced LLMs.

OpenAI Group PBC’s models are also proving to be increasingly adept at finding vulnerabilities. Last month, the company introduced a version of GPT-5.5 that is specifically optimized for cybersecurity research.

Image: Anthropic