Claude Mythos e il Progetto Glasswing: la prima volta nella storia della tecnologia in cui un’azienda ha trattenuto il proprio prodotto migliore per responsabilità. E perché questo cambia le regole per tutti gli altri.
On 7 April 2026, Anthropic announced something without precedent in the history of the technology industry.
A product considered too dangerous to be released to the market.
The model is called Claude Mythos Preview, and it is a general-purpose artificial intelligence that has already identified more than 10,000 critical security vulnerabilities across every major operating system and browser.
According to Anthropic itself, Mythos has reached a level of coding capability that surpasses all but the most highly skilled human experts at finding and exploiting software flaws. And in the wrong hands, it could break into any system.
That is why it is not available. And that is why, in the technology sector, this news is unprecedented.
Project Glasswing: what it is and who is involved
Instead of releasing Mythos to the public, Anthropic has built a controlled consortium. Project Glasswing gives a select group of organisations access to the model, tasked with finding and fixing vulnerabilities before malicious actors can exploit them.
The organisations in the initial consortium include AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks. The second expansion, announced on 3 June 2026, added around 150 new organisations in more than 15 countries, including Okta, Samsung, SK Hynix, SK Telecom, NATO and ENISA, the European cybersecurity agency. The sectors covered include energy, water, healthcare and telecommunications.
The aim is to build a lasting advantage for defenders before Mythos-class AI models become widespread.
Anthropic itself has warned that it will take as little as 6-12 months: within that period, that will be the standard, with the real risk that five, ten or a hundred Mythos models are released without any precautions.
OpenAI has already responded: GPT-5.5-Cyber, its cybersecurity-focused model, has been rolled out to a group of partners for testing.
The first time in the history of technology
Nuclear research, experimental drugs, military technologies. There are precedents for this model of controlled distribution, but never in software.
For the past thirty years, the logic of the technology industry has always run the other way: release as early as possible, iterate in public, grow fast. Mass distribution was the value model. The network only made sense if everyone took part. The most widely distributed product was the product that won.
Anthropic has reversed this logic. It built the most capable product in its history and decided that controlled distribution was preferable to mass distribution. A choice with a measurable economic cost, and one it made anyway.
A deliberate act of governance. Perhaps the most significant a technology company has ever undertaken.
The market implications no one is calculating
If the Glasswing model took hold, the AI market would converge towards a structure closer to energy infrastructure or communications networks than to traditional software.
The most capable models would thus remain accessible only to organisations that meet verifiable security requirements, operate in strategic sectors and have the scale needed to bear the cost of access. The largest, most structured and best-capitalised organisations would obtain the most powerful tools. At that point, there would no longer be any competition.
Over time, the compound effect of this asymmetry would be an inevitable structural divide. Organisations with access to frontier models would accelerate on critical activities: defending their digital infrastructure, identifying competitive vulnerabilities, automating highly complex processes. Organisations without that access would find themselves defending what they have with tools already available to those who want to attack them.
In short, there would be two kinds of organisation.
Those with access to frontier AI.
And those without it.
Governing access
The question Glasswing raises is structural.
Who decides which organisations deserve access to the most capable models? On what criteria is that list built? Who updates it, and how often? Is anyone checking whether the organisations included use the capabilities they receive on the agreed terms?
For now, Anthropic has answered these questions with an internal vetting process, the involvement of the US government and sector-based selection. But it has also stated explicitly that the long-term goal is to make Mythos models more widely available, albeit only after developing stronger safeguards. The “when” is uncertain. The “how” is under construction.
Meanwhile, the market is stratifying. Organisations that meet Glasswing’s access criteria are building expertise in tools the rest of the market does not have. That expertise will not vanish when the models become more widespread: it will simply be a built-in advantage that puts them 12-18 months ahead of everyone else.
The name Project Glasswing is inspired by the glasswing butterfly (Greta oto). It is a double metaphor symbolising (i) the bugs and cybersecurity vulnerabilities that remain invisible without Mythos, and (ii) the idea of bringing rival tech giants together in complete transparency of information.
What should be done now?
The wrong response to this scenario is to wait for frontier models to become accessible to everyone before acting.
The right response is to build, at this stage, the foundations that will make it possible to use those models effectively when they arrive. Organisations that do not yet have clear governance of their digital systems, have not mapped their attack surface, and have not defined which decisions can be delegated to automated systems and with what controls, will find themselves managing powerful capabilities without the architecture needed to do so well.
Glasswing is a process: a way of integrating AI into critical activities with governance, monitoring and reporting. That process is worth as much as the tool. If not more.
The competitive advantage Claude Mythos is delivering goes beyond the idea of a model. It is a method.
And that method, Mythos or no Mythos, must be built now, regardless of access to the model.
New Connections (FAQ)
Will Claude Mythos ever be available to the public?
Anthropic has stated that it intends to make Mythos models more widely available, but only after developing stronger safeguards against malicious use. It has not given a date. In the meantime, Claude Security, a product based on the public Opus 4.8 models, is already available for scanning codebases. Mythos-class capabilities will therefore remain under controlled access for at least the next 12-18 months, according to Anthropic’s own estimates.
What will happen when overly powerful artificial intelligence is released?
Anthropic has already said so explicitly: this is the scenario Project Glasswing seeks to prevent, or at least mitigate. If Mythos-class capabilities proliferate without governance, the cost falls on the entire global digital infrastructure, not just on the organisations directly affected. OpenAI has already responded with GPT-5.5-Cyber. The race between capable models and adequate safeguards is under way. The outcome depends on how many organisations build their defences now, while frontier models are still under a controlled regime.
Should a mid-sized company be concerned about this scenario?
Yes, for two reasons. The first is defensive: as Mythos-class models become more accessible, malicious actors included, the attack surface expands for anyone running digital systems, regardless of size. The second is competitive: the organisations now inside the Glasswing consortium are building skills and processes that give them a structural advantage. The time to lay the foundations of sound AI governance is not when the most powerful models become available. It is now.