What is a zero-day?

A zero-day vulnerability is a security flaw in software that the people responsible for fixing it don't know about yet.

The name comes from the idea that the software vendor has had "zero days" to work on a fix — because they don't know the problem exists. That means there's no patch, no update, and no official defence available at the moment the vulnerability is discovered or exploited.

How a Zero-Day Typically Plays Out

1. Someone finds a flaw in a piece of software — an operating system, a browser, an app. The "someone" is usually one of three types: a researcher, a criminal, or an intelligence agency.

2. They keep it secret. If the finder reports the flaw to the vendor, the clock starts ticking on a fix. If they don't report it — and many don't — it stays a zero-day. The vendor has zero warning.

3. They use it to break into systems quietly. With no patch in existence, there's no public defence. Antivirus tools don't know what to look for. Network monitoring rules haven't been written. The exploit is invisible until someone notices the breach after the fact.

4. Eventually the vendor finds out. Either because someone tells them, or because they detect it being used in the wild — usually after a high-profile breach. The clock now starts. They race to build and ship a fix while the exploit is being actively used against their customers.

5. Once a patch is released, it's no longer technically a zero-day — though every unpatched system remains vulnerable. The window between patch release and full deployment is when most damage actually happens.

Why the Window Matters

The further along that lifecycle a flaw is, the more people have it. A flaw discovered today by a single researcher is dangerous to one set of targets. The same flaw, six months later, is often circulating among multiple criminal groups, several intelligence agencies, and at least one offensive-security marketplace.

By the time it makes the news, it's not a zero-day anymore. It's a patch race.

This is why "responsible disclosure" — the practice of telling the vendor first and giving them a window to fix before publishing — exists as an industry norm. It's also why the most dangerous zero-days are the ones that never get reported and instead get used quietly for years.

Why It Matters for This Week's Story

This is exactly what Anthropic's Claude Mythos Preview was doing autonomously — finding zero-days across every major operating system and browser.

The alarming part isn't just that it found them. It's that it found thousands of them without human steering, and some of them were the most dangerous kind: flaws that could let an attacker escalate from a normal user account to full control of the machine. Anthropic's published numbers from the Mythos preview testing:

• Over 2,000 previously unknown software vulnerabilities found in seven weeks of testing
• 271 zero-days in Firefox alone (per Schneier's reporting)
• A 27-year-old bug in OpenBSD that nobody had spotted in nearly three decades of public source-code review
• A 17-year-old remote-code-execution flaw in FreeBSD
• Over 99% of what Mythos found is still unpatched today

That last number is the one that should keep CISOs up at night. The vulnerabilities exist. The model that found them exists. The patches don't.

Why Anthropic Refused to Ship It Publicly

A tool this good at finding zero-days, in the wrong hands, would be extraordinarily dangerous before defenders had a chance to patch anything. The asymmetry is stark:

• Defenders need to fix every vulnerability. Each one is a separate engineering job, a separate release, a separate deployment cascade through every customer's infrastructure.
• Attackers only need to use one. Any unpatched zero-day, on any system the attacker cares about, is a working exploit.

If Mythos went public, the attack-side flywheel would spin up immediately: criminal groups would point it at every piece of software that runs the world, surface the most exploitable findings, and use them before vendors had any way to know which holes to plug first.

Anthropic chose to keep Mythos restricted, share findings privately with affected vendors, and operate it as what amounts to a privately-held cyber capability rather than a commercial product. Anthropic separately launched Claude Security in public beta — a less-capable defensive tool that helps blue-team analysts respond to incidents — but Mythos itself stayed locked down.

The Policy Knot

This decision is what triggered the White House to consider — for the first time — drafting an executive order to pre-vet frontier AI models before public release. A sharp reversal from the deregulatory stance of the same administration just months earlier. The argument that won: a model with Mythos-class cyber capabilities, in arbitrary hands, would create a national-security incident before regulators could plausibly respond.

The other side of the argument: who decides what's "frontier"? What's the audit trail? What's the appeal process for a researcher who builds something the government decides is too dangerous to ship? These questions don't have settled answers in 2026.

What to Take From This

Three things, in order of certainty:

1. The economics of zero-day discovery just shifted. A model that can find thousands of unknown vulnerabilities in seven weeks doesn't have a precedent. Whoever holds that capability holds something genuinely new in the security landscape.

2. The patching infrastructure isn't built for this volume. The UK National Cyber Security Centre warned this week of a coming "patch wave" — a flood of critical updates simultaneously across every layer of the tech stack. Most organizations' patching cadences were built for a world where vulnerabilities arrived a few at a time.

3. AI offensive capability is now a regulatory variable. A year ago, AI safety arguments were mostly about model behaviour — sycophancy, bias, harmful outputs. As of this week, they include "this model will find more zero-days than any defender can patch." That's a different category of risk, and the policy response is still being written.

Worth knowing what a zero-day actually is, before reading the next round of headlines about who's finding them and what's getting locked down as a result.