SAN FRANCISCO, June 2 (Reuters) - Anthropic on Tuesday said it was planning a wider rollout of its artificial intelligence that has caught the attention of government and financial institutions around the world for its ability to find software vulnerabilities.

Project Glasswing, Anthropic's partnership of around 50 organizations using its AI to secure software, is expanding to roughly 200, the San Francisco-based startup said. Each member must meet certain security requirements before accessing Anthropic's powerful AI called Claude Mythos Preview, it said.

Anthropic collaborated with the U.S. government, existing Glasswing partners and others in the security industry on the rollout, it said.

The news comes a day after Anthropic filed confidentially for what could be one of the biggest initial public offerings ever, culminating a frenzy of business and government interest in its AI.

Its April Mythos Preview announcement sowed panic that AI could hack software that was not first buttoned up, thanks to Anthropic's technology, which can surface ways to stitch together and exploit small bugs.

U.S. Treasury Secretary Scott Bessent and then-Federal Reserve ​Chair Jerome Powell convened an urgent meeting with bank CEOs to warn of ‌the risks, while financial institutions pushed for access and raced to plug holes.

Anthropic said its partners have found more than 10,000 security flaws they considered highly or critically severe, though some industry experts have said fears of unfettered hacking have been overstated.

GOLDMAN SACHS: MORE INFORMATION, GUARDRAILS KEY

Goldman Sachs CEO David Solomon said at an economic event on Tuesday that information about these models should be shared more broadly among companies, while emphasizing the need for guardrails, as open-source models can quickly do the same things.

"They're very, very powerful models," he said.

"What I can say is that a large group of companies are doing a lot of testing. It's also important that the information is shared more broadly so other companies and the economy can learn."

Goldman is among the banks testing Mythos.

Anthropic declined to name the organizations it is bringing into the fold.

However, it said its new Glasswing partners are based in more than 15 countries and boost defenses in the healthcare, power, water, communications and hardware industries. The expansion includes government organizations, an Anthropic spokesperson said.

A big attack on partner organizations could affect more than 100 million people, Anthropic estimated. Project Glasswing aims to bolster such software infrastructure before AI this powerful is widely available from other providers, a prospect that is no more than six to 12 months away, Anthropic said.

The startup has also said it aims to bring Mythos-class models to all its customers, with additional safeguards, in the coming weeks.