Anthropic Expands Restricted AI Preview to Critical Infrastructure Sectors
The company cites catastrophic security risks and strict access requirements as it delays public release of the frontier AI model
Anthropic has expanded its Project Glasswing initiative, inviting approximately 150 additional organisations to preview the Claude Mythos artificial intelligence model. The new cohort, drawn from more than 15 countries, includes entities from sectors such as public utilities and healthcare, broadening the scope of the programme beyond its original technology-focused group.
The expansion follows the initial launch in April, which saw nearly 50 partners granted access. That first group comprised major technology firms including Apple, Amazon, Broadcom, CrowdStrike, Microsoft, and NVIDIA. The new invitees represent industries that were not well-represented in the original cohort, specifically power, water, and telecommunications providers, as well as healthcare organisations.
Access to the model remains strictly controlled. Anthropic requires all participants to meet rigorous security standards before being granted entry. The company emphasised that the selected partners share a common vulnerability profile, noting that a successful cyber attack on their codebases could have catastrophic consequences.
Anthropic estimates that a major breach affecting most partners could impact more than 100 million people, carrying significant ramifications for both global and national security. The company stated it is working to release the model publicly only after implementing highly robust safeguards to prevent the misuse of its advanced coding and vulnerability exploitation capabilities.
The development of Claude Mythos has already influenced regulatory discussions. In May, the announcement appeared to prompt the Trump administration to consider an executive order for sharing vulnerability information with critical institutions. However, the signing of that order was reportedly postponed due to objections from the administration and the tech industry.
Anthropic maintains that it is working as quickly as possible to safely release Mythos-level capabilities to the wider public. The company reiterated that it will not make the model widely available until it is confident that appropriate measures are in place to mitigate potential risks associated with its cyber capabilities.
The programme was established because Anthropic observed that AI models had reached a level of coding capability capable of surpassing most humans in identifying and exploiting software vulnerabilities. By limiting access to vetted partners, the company aims to manage these risks while testing the model’s impact on critical infrastructure security.
