Anthropic has unveiled a preview of Mythos, its most capable model to date, deploying it through a controlled security initiative called Project Glasswing. Twelve partner organisations, including Amazon, Apple, Microsoft, and Cisco, are using the model to find vulnerabilities in critical software systems.
The company says Mythos has already identified thousands of zero-day vulnerabilities, some of them decades old. That is a striking claim, and one that raises an obvious question: if the model is good enough to find those flaws, it is good enough to exploit them.
Anthropic acknowledged as much. In a previously leaked internal document, the company warned that Mythos could pose a serious cybersecurity threat if weaponised by bad actors. It is a tension the AI industry has never fully resolved, and Mythos makes it sharper.
A model above Opus
Until now, Anthropic's Opus series sat at the top of its model hierarchy. Mythos sits above it. The leaked document, first reported by Fortune after an internal blog post was left in an unsecured data cache, described the model as a new tier entirely, larger and more capable than anything the company has previously released.
That description suggests this is more than an incremental upgrade. The company's own framing, calling it "by far the most powerful AI model we've ever developed," is unambiguous. Whether real-world performance matches that billing will only become clear as the Project Glasswing partners begin to report their findings.
Security theatre, or something more?
The controlled rollout is deliberate. Only 12 organisations are classed as official partners, though 40 in total will gain access to the preview. Anthropic says those partners will share what they learn, with the broader tech industry expected to benefit from the knowledge gathered.
That framing serves a dual purpose. It limits exposure of a model Anthropic itself acknowledges could be dangerous in the wrong hands. It also builds a coalition of credibility, drawing in household names from across the technology sector ahead of what is likely to be the most scrutinised IPO of the decade.
The IPO context
Anthropic is targeting a Nasdaq listing as early as October 2026. Bankers expect the raise to exceed $60 billion, according to The Information, which would make it one of the largest public offerings in history.
Annualised revenue reached $19 billion by March 2026, up from $9 billion at the end of 2025. TECHi® Enterprise customers account for 80% of that figure. Eight of the Fortune 10 are clients. Claude Code alone generates $2.5 billion in annualised revenue and holds a 54% share of the AI coding tools market. TECHi®
Against that backdrop, Mythos is not just a technical announcement. It is a signal to institutional investors that Anthropic's model development pipeline has not plateaued. The company is spending heavily, with planned outlay on model training and inference running at roughly $19 billion in 2026, approximately matching its revenue. TECHi® It needs to demonstrate that the capability gains justify the cost.
The legal complications
There is another layer of complexity. Anthropic says it has been in discussions with federal officials about Mythos. Those conversations are likely strained. The Pentagon has labelled Anthropic a supply-chain risk, and the company is currently in legal dispute with the Trump administration after refusing to allow autonomous targeting or surveillance of US citizens.
That stand earned Anthropic credibility in some quarters and hostility in others. For a company moving toward a public listing, the legal uncertainty is a risk factor that will require careful disclosure.
Breakthrough or incremental?
The honest answer is that it is probably both, depending on the frame. As a product launch, Mythos is tightly controlled and narrow in scope. It is not available to the public. Its deployment is limited to a single use case.
Related reading
- UK government courts Anthropic with London office and dual listing offer
- Anthropic accidentally leaked the blueprint for its $2.5bn AI coding product. Here is what it reveals
- AI software disruption story is wrong and Anthropic's own demos prove it
As a statement of where Anthropic's research has reached, it is more significant. A model that can find critical vulnerabilities in software that has gone unpatched for 20 years is operating at a level of capability that most cybersecurity firms cannot match with human analysts alone.
The real test will come when the Project Glasswing partners publish their findings. If the zero-day claim holds up at scale, the debate will shift quickly from whether Mythos represents a breakthrough to whether models this capable should ever be released at all.
