maximum defensive impact through shared findings
In a quiet but consequential shift, Anthropic has chosen to open the channels of knowledge that its tightly controlled Glasswing cybersecurity program once kept sealed. Where caution once governed the sharing of discovered vulnerabilities, the company now believes that the greater good lies in transparency — allowing its partners among the world's largest technology firms to carry hard-won defensive insights outward, into the broader ecosystem of institutions and communities that depend on secure software. It is a recognition, perhaps, that in the long contest between those who would harm and those who would protect, the defenders are strongest when they speak to one another.
- Mythos, Anthropic's AI model capable of both finding and theoretically exploiting software vulnerabilities, was always a tool powerful enough to demand careful handling — and that power made secrecy feel necessary at the start.
- Participating companies like Amazon, Microsoft, Nvidia, and Apple had written confidentiality clauses into their own contracts, wary that disclosing what they found might paint targets on their backs.
- As weeks passed and the program matured, Anthropic concluded that locking findings inside Glasswing was costing the wider world more than it was protecting any single partner.
- Partners may now share vulnerabilities, defensive tools, and code with security teams, regulators, open-source maintainers, journalists, and the public — so long as responsible-disclosure norms are honored.
- The Pentagon is already running Mythos against U.S. government systems to patch weaknesses, even as it plans to reduce its dependence on Anthropic — a signal that the tool has crossed from experiment into operational reality.
Anthropic announced this week that it is easing the confidentiality restrictions governing Project Glasswing, its controlled cybersecurity initiative involving Amazon, Microsoft, Nvidia, and Apple. Partners in the program, which tests the unreleased Mythos model for strictly defensive purposes, can now share their findings — including discovered vulnerabilities, defensive tools, and developed code — with organizations well beyond the program's original boundaries.
Mythos, introduced in early April, marked a notable advance in what AI can accomplish in cybersecurity. Its sophisticated coding capabilities allow it to identify weaknesses in software and, in principle, devise ways to exploit them — a dual-use potential that led Anthropic to structure Glasswing as a tightly restricted initiative from the outset. The confidentiality clauses were not imposed by Anthropic but written into individual contracts at the partners' own request, as companies sought assurances before sharing sensitive findings.
Over the weeks since the program launched, Anthropic reconsidered that approach. The company now holds that the defensive value of open information-sharing outweighs the original concerns. Partners may disclose findings to security teams at other firms, industry groups, government regulators, open-source maintainers, journalists, and the public — provided they follow responsible-disclosure practices, giving affected parties time to issue patches before any public announcement.
The policy shift arrives as the Pentagon is already deploying Mythos to identify and remediate vulnerabilities across U.S. government systems, even while working to reduce its long-term reliance on Anthropic. Together, the expanded sharing rules and active government deployment suggest that Mythos is transitioning swiftly from a carefully guarded experiment into a piece of operational security infrastructure — with defensive knowledge now moving more freely across industry and government alike.
Anthropic announced this week that it is loosening the confidentiality restrictions it had originally placed on a select group of technology companies testing its Mythos cybersecurity model. The shift means that partners in what the company calls Project Glasswing—a controlled program that includes Amazon, Microsoft, Nvidia, and Apple—can now openly discuss their involvement and share the vulnerabilities they discover, along with defensive tools and code they develop, with organizations outside the program.
Mythos, which Anthropic introduced in early April, represents a significant leap in what artificial intelligence can do in the cybersecurity space. The model's ability to write sophisticated code gives it an unusual capacity to both identify security weaknesses in software and, theoretically, devise methods to exploit them. This dual capability is precisely why Anthropic structured Glasswing as a tightly controlled initiative in the first place, granting access only to major technology firms willing to use the unreleased Claude Mythos Preview model strictly for defensive purposes.
When the program began, participating companies asked for confidentiality protections. They wanted assurances before sharing sensitive findings about vulnerabilities—partly out of concern that disclosing such information might make them targets for attackers. Anthropic built those protections into the agreements each partner signed. But as the program has matured over the past several weeks, the company has reconsidered that approach. An Anthropic spokesperson explained that while there was never a formal nondisclosure agreement governing Glasswing, the confidentiality clauses were written into individual partner contracts at the companies' request.
The company now believes that the defensive benefit of sharing these findings outweighs the original security concerns. Partners are now permitted to disclose their findings to security teams at other companies, industry groups, government regulators, open-source software maintainers, journalists, and the general public—provided they follow responsible-disclosure practices, which typically means giving affected parties time to patch vulnerabilities before public announcement. "We fully support our partners sharing findings with each other and companies outside of Glasswing to triage vulnerabilities," the Anthropic spokesperson said. "As the program has matured, we've adapted them to ensure key information can be shared broadly—including outside the program—for maximum defensive impact."
The timing of this policy shift is notable given that the Pentagon has already begun deploying Mythos to identify and fix software vulnerabilities across U.S. government systems. According to the Defense Department's top technology official, the military is using the model even as it works to transition away from relying on Anthropic—a move that suggests the government sees real value in the tool despite broader concerns about AI dependency. The combination of Glasswing's expanded information-sharing rules and the Pentagon's active deployment suggests that Mythos is moving rapidly from experimental program to operational security infrastructure, with findings and defensive strategies now flowing more freely across the technology industry and into government agencies.
Citas Notables
We fully support our partners sharing findings with each other and companies outside of Glasswing to triage vulnerabilities.— Anthropic spokesperson
As the program has matured, we've adapted them to ensure key information can be shared broadly—including outside the program—for maximum defensive impact.— Anthropic spokesperson
La Conversación del Hearth Otra perspectiva de la historia
Why did Anthropic originally lock down information sharing if the whole point was to find vulnerabilities faster?
The companies in the program were nervous. They wanted to test a powerful tool without broadcasting every flaw they found. If word got out that Amazon or Microsoft had a vulnerability, attackers would target them before patches were ready.
But now they're opening it up. What changed?
The program proved itself. After a few weeks, Anthropic realized that keeping findings secret actually slowed down the broader defense. If Microsoft finds a vulnerability pattern, Apple needs to know about it too. The industry moves faster when information flows.
So this is about speed over secrecy.
Exactly. And there's a calculation underneath: a vulnerability that stays hidden is a vulnerability that might be exploited. One that's shared responsibly—given time for patches—protects everyone.
What about the Pentagon using this? That seems like a different animal entirely.
It is. The government is already running Mythos on its own systems, finding its own problems. That's not theoretical anymore. It's operational. The question now is whether the military can actually reduce its dependence on Anthropic while still using the tool.