Project Glasswing has a partner list, a model name, and a budgeted promise. It still does not have public access.

The paper's May 17 brief on Glasswing Day 40 asked for a public artifact, not mythology. Monday's source page gives plenty of institutional detail but preserves the same asymmetry.

Anthropic says Project Glasswing brings together Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks to secure critical software. It says launch partners will use Claude Mythos Preview in defensive security work. [1]

The page also says Anthropic extended access to more than 40 additional organizations that build or maintain critical software infrastructure, committed up to $100 million in usage credits, and pledged $4 million in direct donations to open-source security organizations. [1]

This is not nothing. It is also not a public release. The useful distinction is between an access regime and an open tool. Partners may scan code, learn from vulnerabilities, and share some findings, while outsiders still cannot test the system, audit selection, or see the conditions under which access is granted.

MSM can write the initiative as a cybersecurity advance. X can write it as a cartel. The paper's middle is less satisfying and more accurate: if frontier models can find serious flaws faster than humans, the access list becomes governance.

That is why Daybreak made Glasswing harder to describe. OpenAI turned trusted AI access into a product page. Anthropic has a powerful page too, but its public record still points inward: named partners, selected additional organizations, and defensive work behind the gate.

The comparison now lives in the company pages themselves. OpenAI has Daybreak and Anthropic has Glasswing, and both point toward the same pressure question: if the most capable systems are being allocated through trusted lists, who decides what trust means?

Anthropic can answer that some constraint is prudent. A model capable enough to find important vulnerabilities may also be capable enough to help bad actors if access is casual. The company's page leans hard into defense, critical software, and support for the open-source ecosystem. [1] Those are serious claims, and the partner list includes organizations that actually maintain infrastructure people rely on.

But prudence does not remove the governance problem. A closed gate may be safer than a public free-for-all, yet it can still concentrate advantage among the largest platforms, banks, cloud companies, and security vendors. Open-source maintainers outside the chosen circle may receive donated support without receiving equal visibility into the tool that found the flaws. That is a distribution question, not an anti-security slogan.

The next receipt is not another launch essay. It is a public charter, selection criteria, vulnerability-disclosure data, or a route for qualified outsiders that makes the gate legible. Until then, Glasswing is best described as a partner program with public-interest promises, not as public access.

-- DAVID CHEN, Beijing