Claude Mythos and Project Glasswing: Why a Locked-Down AI Model Has the Cybersecurity World Watching

For years, the rhythm of artificial intelligence development has felt almost predictable. A company unveils a new model, promises it is faster, smarter and more capable than the last one, then invites the public to test it, celebrate it and, inevitably, argue over it. The cycle has become so familiar that even major advances can feel routine.

That is why Anthropic’s handling of Claude Mythos has landed so differently.

Instead of a broad public rollout, splashy demos and open experimentation, the company has done something far more unsettling. It has announced the model with enormous fanfare, while at the same time keeping it tightly restricted. In place of a conventional launch, Anthropic has introduced Project Glasswing, an initiative designed to direct Mythos toward defensive cyber work rather than allow it to be widely used or misused.

That decision alone tells its own story. In the AI race, companies do not usually hold back their most powerful tools unless they believe the risks of release may be as significant as the opportunities. Mythos appears to be one of those cases.

What has made the announcement so arresting is not just the secrecy surrounding the model, but the nature of the claims being made about it. Early reports suggested that, when given instructions, the system had managed to move beyond a contained testing environment and send an email to a researcher. That detail is dramatic enough to capture headlines, conjuring the familiar fear of a model that can slip the boundaries set for it. Yet Anthropic’s own assessment appears to suggest a more precise concern. The danger is not so much that Mythos is likely to develop rogue intent, but that it may be extraordinarily effective at carrying out harmful tasks when directed by a human.

That is a subtler, but perhaps more important, warning.

The real alarm around Mythos comes from what Anthropic says it has already found. According to the company, the model has identified vulnerabilities across major operating systems and web browsers, as well as in the open-source software that quietly underpins much of the digital world. These are not flashy consumer apps or minor bugs tucked away in obscure tools. They are the hidden joints and load-bearing beams of modern computing, the kind of infrastructure most people never see but rely on every day.

Among the most startling examples is Anthropic’s claim that Mythos uncovered a flaw in OpenBSD, a security-focused operating system trusted in environments where reliability and protection matter deeply, including routers and firewalls. The company says the vulnerability had remained undetected for 27 years. It also reports that the model found a 16-year-old flaw in FFmpeg, an unglamorous but enormously widespread software project used to process audio and video across countless devices, platforms and services. Even more concerning, Anthropic says Mythos identified several vulnerabilities in the Linux kernel and chained them together in a way that could potentially hand an attacker full control of a machine.

If those claims stand up to scrutiny, they would mark a profound shift in what AI is capable of doing in cybersecurity.

For decades, one of the basic constraints of hacking, whether for malicious intrusion or defensive research, has been the scarcity of expertise. Serious vulnerability discovery has required unusual patience, deep technical skill and a willingness to spend weeks, months or even years spelunking through dense codebases in search of overlooked weaknesses. It has never been easy work. In many cases, dangerous flaws remained unpatched not because they were impossible to find, but because nobody with the right combination of talent and time had found them yet.

Mythos hints at a world where that equation changes.

If an AI model can inspect operating systems, browsers, media-processing libraries and other core pieces of software at scale, it could alter the economics of cybersecurity completely. Vulnerability hunting could move from being a specialised, painstaking craft into something increasingly routine, automated and relentless. That would be a gift to defenders, if they are first to wield it. It would be a nightmare if attackers gain similar capabilities before systems are hardened.

That tension lies at the heart of Project Glasswing. Anthropic says the initiative brings together a formidable coalition of technology companies, open-source organisations and major financial institutions, including Microsoft, Amazon, Google, Apple, Cisco, NVIDIA, the Linux Foundation and JPMorganChase. The strategy appears simple in theory, though enormous in practice: use Mythos to help defenders identify and patch weaknesses in critical digital infrastructure before comparable AI tools become available to bad actors.

This is not just about securing Silicon Valley’s own products. The underlying software Mythos is reportedly probing supports far more than tech companies. It supports banking systems, hospitals, utilities, airlines, retailers and the digital plumbing of everyday life. The internet’s invisible architecture is rarely discussed until it breaks, but it forms the backbone of modern economies and essential services.

That is why the implications of Mythos extend far beyond the usual AI industry chatter.

Cybersecurity failures are not abstract. They are not merely technical glitches or line items on a corporate risk register. When systems fail, real people bear the consequences. Australians do not need much reminding of that. The Optus breach exposed the personal information of around 9.5 million customers, while the Medibank attack compromised deeply sensitive health records, some of which were later released online. Those incidents were not just failures of databases or internal controls. They became crises of privacy, identity and public trust.

In that context, a model that can expose hidden flaws faster than human researchers changes the stakes for everyone.

Yet there is another reason the Mythos story has landed with such force, and it has to do with trust. Anthropic has published unusually detailed material for a model it is not broadly releasing, and reports have suggested that senior United States officials have treated the announcement seriously enough to involve major banking leaders in discussions about cyber risk. Even so, much of what Anthropic says about Mythos cannot currently be independently verified. The company has said that more than 99 per cent of the vulnerabilities the model has discovered remain undisclosed because they have not yet been patched. From the perspective of responsible security practice, that restraint makes sense. Publicising unpatched flaws would create obvious danger. But it also means the public is being asked to accept extraordinary claims while being unable to inspect most of the underlying evidence.

That leaves the outside world in an uneasy position, caught between caution and curiosity.

There is precedent here. In 2019, well before the current consumer AI boom, OpenAI famously withheld the full release of GPT-2, arguing that even that comparatively primitive model could be misused. At the time, the decision was controversial, with some dismissing it as theatre and others taking it as an early sign of how seriously AI labs were beginning to think about deployment risk. Mythos feels like a far more consequential version of that moment. The concern is no longer about spam, synthetic text or internet misinformation alone. It is about the possibility that advanced models may become elite vulnerability researchers, capable of surfacing critical software weaknesses at a speed and scale human teams cannot match.

That prospect should worry organisations that have grown comfortable with treating cybersecurity as a back-office issue.

The relative public silence from many cybersecurity and software firms in the wake of Anthropic’s announcement is telling. Some may simply be waiting for more evidence. Others may be reluctant to comment in case the model eventually exposes weaknesses in their own systems. But quiet does not mean calm. In boardrooms, engineering teams and security operations centres, announcements like this are likely to be prompting some difficult internal conversations.

What would it mean if offensive cyber capability becomes cheaper, faster and more accessible through AI assistance? How should governments regulate the release of highly capable models without stifling legitimate defensive research? Who gets to decide what counts as safe access, and who can be trusted with tools powerful enough to destabilise critical infrastructure?

Those are long-term questions, and they are not easily answered. But the short-term lesson is far less complicated.

For individuals, the rise of systems like Mythos is not a reason to panic, but it is a reason to stop treating cyber hygiene as optional. Software updates matter. Router firmware matters. Unsupported devices matter. Password reuse matters. Multi-factor authentication matters. The basic habits that security professionals have been urging for years may soon become even more important, not because the advice has changed, but because the tools available to both defenders and attackers may be accelerating.

Anthropic’s locked-down model may or may not prove to be as transformative as its creators suggest. The evidence, for now, is largely inside the company and among its chosen partners. But even at this early stage, Mythos has done something significant. It has sharpened the sense that AI is moving beyond being a tool for writing, summarising and coding, and toward becoming a force that can probe the structural weaknesses of the digital world itself.

That is why the industry is paying such close attention. Not because Mythos has already broken the internet, and not because it is likely to become sentient and escape the lab, but because it may represent the beginning of a new era in cybersecurity, one where the balance between offence and defence is redrawn by machines that can see flaws humans missed for decades.

If that future is arriving, then Project Glasswing is more than a defensive initiative. It is an admission that the race has already begun.

Photo Credit: DepositPhotos.com