Zürcher Nachrichten - Anthropic expands access to powerful Mythos AI model

The California AI company had initially restricted access to Mythos in early April, after tests showed the model was extraordinarily good at finding security holes in software code -- identifying thousands of them in a short period of time.

At first, only around 50 partners -- mostly US companies -- were brought into the program called Project Glasswing to test the model's powers before it could be weaponized by hackers.

Among them were major tech companies including Amazon, Google, Nvidia, Apple and Microsoft, along with cybersecurity firms CrowdStrike and Palo Alto Networks.

By late May, those early partners had already used Mythos to uncover more than 10,000 serious security weaknesses that hackers could exploit to cause real damage.

Now Anthropic is opening the program to groups from more than 15 countries.

Most of them run software that millions of other organizations -- including governments -- depend on every day.

A successful cyberattack on any one of them, Anthropic says, could affect "more than 100 million people, with important ramifications for both global and national security."

The new participants cover industries that were missing from the original group, including power grids, water systems, hospitals, phone networks and hardware manufacturers.

The company declined to name the new partners.

When Project Glasswing launched, some critics accused Anthropic of overhyping the threat to generate attention.

The company, which is also behind the popular Claude chatbot, announced on Monday that it had filed confidentially for an initial public offering (IPO).

But companies that have tested Mythos have since backed up its capabilities, with the US government, which had been in a legal dispute with Anthropic, eventually testing the model over security concerns.

Some observers link that government shift to a broader policy change from the Trump administration, which struck deals with Google, Microsoft and Elon Musk's xAI in early May to have their most powerful AI models reviewed for safety before going on sale.

A presidential order to make that policy official was drafted but then put on hold after disagreements inside the White House.

On Monday, a European Commission spokesperson confirmed to AFP that Anthropic had invited ENISA -- the European Union's cybersecurity agency -- to join Project Glasswing.

Y.Keller--NZN