Anthropic Won't Release Mythos — What That Means for You

The first AI model Anthropic is afraid to ship

On April 7, Anthropic introduced Claude Mythos Preview — and announced in the same breath that it will not be made publicly available. Mythos found a 27-year-old bug in OpenBSD, a 16-year-old flaw in FFmpeg, and thousands of other vulnerabilities across every major operating system and web browser in testing. Over 99% of what it found is still unpatched. The company is limiting access to roughly 40 critical-infrastructure partners through a new initiative called Project Glasswing rather than releasing it broadly.

This is the first time Anthropic — or any frontier AI lab — has withheld a model from public release specifically because of its offensive cybersecurity potential. For small businesses, the question is not whether Mythos-class capability will eventually reach the internet. The question is what you do between now and then.

What actually happened

Mythos Preview is a general-purpose Claude model with one specific talent turned up to unusual levels: finding and exploiting software vulnerabilities without human help. In Anthropic’s internal tests, Mythos wrote a working web browser exploit that chained four separate bugs into a complete sandbox escape. Against Firefox, the previous-generation Claude Opus 4.6 managed working exploits twice out of hundreds of attempts — Mythos produced 181 working exploits plus partial control on 29 more.

The upgrade is not incremental. It is a step change.

Partners getting early access include Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks, among others. The stated goal: use Mythos to harden the world’s most critical software before comparable capability is available more widely.

The problem Anthropic is openly worried about is asymmetry. As security firm RunSafe’s CTO Shane Fry told Fortune this week, “vulnerability discovery is outpacing patching.” Remedio founder Tal Kollender put it more bluntly: defenders face “a race they’re not yet equipped to win” for roughly the next year.

Why this matters for small businesses

You do not run Mythos. You do not get Project Glasswing access. None of that makes you safe.

Here is what the next 12 months look like for a small business running Windows laptops, a point-of-sale system, a couple of cloud SaaS accounts, and a website:

Your software stack is the discovery target. When Mythos finds a zero-day in Windows, macOS, Chrome, or a common CMS, that flaw exists in your environment the moment it is discovered. The window between discovery and patch is the window you are exposed.
That window is getting longer, not shorter. Anthropic’s own data shows 99% of Mythos-found vulnerabilities are still unpatched because disclosure follows a 90+45 day responsible-disclosure timeline. Even under ideal conditions, vendors take months to ship fixes.
Attackers will get their own Mythos. Anthropic is restricting its model. Open-source labs, adversarial nation-states, and criminal groups are building their own. The Bloomberg feature “How Anthropic Discovered Mythos AI Was Too Dangerous For Release” makes clear the capability will escape the controlled-partner pool eventually.

Small businesses are already taking the brunt of AI-assisted attacks. The Identity Theft Resource Center found that 81% of small businesses suffered a breach in 2025, with AI implicated in more than 40% of those incidents. Mythos-class tools will widen that gap, not close it.

Our take

The defensive AI story has been real for a while. We have written before that cybersecurity is the AI use case where small businesses get the clearest ROI — and it still is. What Mythos changes is the pressure. Before this week, “I’ll get around to patching” was a risk. Now it is a countdown.

The bottom line: The gap between when a vulnerability is found and when you can realistically install the patch is about to become the single most important security metric you track.

A few things are missing from most coverage of the Mythos announcement. First, the restrictions are voluntary and unenforceable long-term — other labs will ship similar models under different names. Second, the tool cuts both ways: the same capability that finds zero-days can also help small businesses scan their own codebases, SaaS configurations, and third-party dependencies. Anthropic’s own guidance is to start using current frontier models for defensive work now, not to wait.

The open questions are regulatory and contractual. Will cyber insurance carriers start requiring AI-assisted patching programs? Will software vendors face new liability for slow patch rollouts? Appalachian small businesses — especially those serving healthcare, financial, or government customers — should expect their downstream contracts to start spelling out patch-window expectations within the next year.

What you should do this month

You do not need a security team to take meaningful action. Four moves in order of impact:

Turn on auto-update everywhere you can. Operating systems, browsers, phones, routers, and any CMS or e-commerce platform you run. This is free, takes an afternoon, and closes most of the window for opportunistic attacks.
Inventory your software. You cannot patch what you do not know you run. Write down every piece of software and every SaaS tool your business touches — including the ones one person uses for one thing. Note who owns each one and how it updates.
Set a 30-day patch review. Once a month, check the tools that do not auto-update. Point-of-sale firmware, specialty industry software, and anything on-premise tends to lag. A recurring calendar reminder beats a written policy nobody reads.
Write down what happens if something goes wrong. A one-page incident response plan — who to call, how to take systems offline, where backups live — is the single highest-leverage document in small business security. SBA’s cybersecurity resources and CISA’s small business guide both include free templates.

Signals worth watching over the next few months: which peer labs announce Mythos-equivalent models, whether any are released open-weight, and how quickly patch SLAs in your vendor contracts tighten. These are the leading indicators for how quickly Mythos-class risk arrives on your doorstep.

The countdown has started

Anthropic’s decision to restrict Mythos is a rare piece of honest corporate risk-telling — the AI industry is moving faster than the patch infrastructure the world is built on. The good news is that most of what small businesses should do in response is not expensive or complicated. It is consistent. Auto-updates, an inventory, a patch cadence, and a plan for the day something breaks.

Working through how to harden your business against AI-assisted threats? Get in touch — we help small businesses build practical AI and security programs without enterprise budgets.