More than two dozen companies including JPMorgan Chase & Co. and an array of cybersecurity firms are collaborating to remedy software flaws spotted by cutting-edge artificial intelligence models.

The coalition, called Athena and led by cybersecurity startup Chainguard, involves Cisco Systems Inc., Cloudflare Inc. and others, Chainguard said Monday. The group is focused on securing open-source software, which underpins technology used across the economy, from web browsers and data centers to smartphones and ATM machines.

In an interview, Chainguard Chief Executive Officer Dan Lorenc said members of the group are searching for potential vulnerabilities using advanced AI models, such as OpenAI’s GPT-5.5-Cyber.

Anthropic PBC said Friday it would disable access to its most advanced artificial intelligence models, including Mythos, following an unprecedented order by the Trump administration to keep the technology out of the hands of all foreign nationals. Access to GPT-5.5-Cyber is currently limited to those approved by the AI startup due to concerns about their capabilities and potential for harm.

Anthropic and rival OpenAI have spent much of the past year one-upping each other with a flurry of AI model releases aimed at streamlining a wider range of professional tasks — from coding to financial services to healthcare — with the goal of courting more business customers and justifying their lofty valuations. In recent months, AI’s potential utility in cybersecurity has emerged as a focal point for both companies.

So far, the coalition said it has examined over 20,000 potential vulnerabilities shared by members — some of which turned out to be duplicates — and sent out more than 2,000 patches for 500 open-source projects.

“We know the timeline on all of this is very limited,” said Lorenc, whose company focuses on securing open-source software. That’s because while Anthropic and OpenAI have limited access to their most capable cybersecurity-related AI models, he suspects that open-source AI models will eventually be just as good, providing attackers access to the same kinds of tools.

Most of the vulnerabilities so far have been found using Anthropic’s Mythos AI model, Lorenc said, and a growing number from GPT-5.5-Cyber. He said they include “very serious software vulnerabilities,” but declined to provide more details ahead of a series of disclosures of the issues that will start next month.