SAN FRANCISCO, June 2 - Anthropic announced on Tuesday that it plans to expand Project Glasswing, its collaborative effort that uses the company's AI to identify software vulnerabilities, from about 50 participating organizations to roughly 200.
The San Francisco-based startup said each organization joining the expanded Glasswing network must satisfy specific security requirements before receiving access to Anthropic's Claude Mythos Preview model. Anthropic said it worked with the U.S. government, current Glasswing participants and other firms in the security sector during the rollout of the program.
The announcement arrives one day after Anthropic filed confidentially for an initial public offering. The company has drawn intense attention from both government bodies and financial institutions following its April release of the Mythos Preview, which demonstrated the model's ability to surface ways to combine small software bugs and potentially exploit them.
Anthropic said its Glasswing partners have identified more than 10,000 security flaws they classified as highly or critically severe. The company declined to name the organizations being added to the program, but said the new partners operate across more than 15 countries and span industries including healthcare, power, water, communications and hardware. Anthropic also confirmed that government organizations are among the additions.
Financial and regulatory concern escalated after the Mythos Preview disclosure. The announcement prompted warnings that the model could reveal attack paths in software that had not been fully secured. U.S. Treasury Secretary Scott Bessent and then-Federal Reserve Chair Jerome Powell convened an urgent meeting with bank chief executives to highlight those risks, and financial institutions sought access to the model while working to plug vulnerabilities.
Goldman Sachs is among the banks testing Mythos, and Goldman CEO David Solomon said at an economic event on Tuesday that companies should share information about the capabilities of these models more broadly - provided there are appropriate guardrails in place. "Theyre very, very powerful models," he said, noting the importance of sharing information so other companies and the broader economy can learn from testing efforts.
Anthropic estimated that a large-scale attack on partner organizations could affect more than 100 million people. Project Glasswing is portrayed by the company as a way to strengthen critical software infrastructure ahead of the time when comparable AI capabilities could be more broadly available from other providers, a prospect Anthropic estimated at six to 12 months away.
The startup also said it plans to make Mythos-class models available to all its customers in the coming weeks, accompanied by additional safeguards.
The company emphasized that access to Claude Mythos Preview will be controlled and limited to organizations that meet its security criteria. Anthropic has said it collaborated with existing Glasswing partners and security professionals to shape the expansion and the guardrails that will govern use of the model.
Beyond the immediate security implications, the announcement has drawn attention from market participants. A promotional note included in coverage of the corporate developments highlighted that ProPicks AI evaluates Goldman Sachs alongside thousands of other companies monthly using more than 100 financial metrics, and that the tool has identified notable past winners such as Super Micro Computer and AppLovin, with cited returns of +185% and +157% respectively. That commentary positions investor interest in financial firms that are testing or seeking access to Mythos-class models.
Context and next steps
Anthropic's expansion of Project Glasswing increases the number of vetted organizations using Claude Mythos Preview to roughly 200, with those groups required to meet predetermined security standards. The company continues to collaborate with government entities and industry partners and plans broader customer access to Mythos-class models with extra safeguards in the near term.