AI can now find serious software flaws faster than human researchers can. Anthropic says its Mythos Preview model, which it does not plan to make generally available, has already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser. For a small business, that means one thing: the time between a flaw being found and someone using it against you is getting shorter. Patch your internet-facing equipment faster, and scan your own systems (with permission) before attackers do.

This post covers what was announced and why it matters to a clinic, law office, accounting firm, contractor or restaurant group that doesn't have a security team. Then it gives you a patch schedule and a scanning plan you can actually stick to.

What Anthropic announced, and why it kept the model locked up

On April 7, 2026, Anthropic announced Project Glasswing. It brings together Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Anthropic has also given access to over 40 additional organizations that build or maintain critical software infrastructure. Infosecurity Magazine describes the effort as an initiative that uses AI to identify and remediate undiscovered vulnerabilities in critical software.

The examples are the alarming part:

Anthropic's technical write-up says the model can identify and then exploit zero-day vulnerabilities in every major operating system and every major web browser when a user directs it to. It also says over 99% of the vulnerabilities it has found have not yet been patched.

Anthropic says it does not plan to make Mythos Preview generally available. The Hacker News reports that the decision was driven by the model's cybersecurity capabilities and concerns they could be abused. Instead, Anthropic is committing up to $100M in usage credits for Mythos Preview and $4M in direct donations to open-source security organizations.

Treat these counts and examples as the vendor's own claims. The direction is clear anyway, and it matches what defenders have been seeing for years.

1000s
High-severity bugs Anthropic says Mythos Preview found
27 yrs
Age of the OpenBSD bug it found
99%+
Of the bugs found, share not yet patched

AI finds bugs faster, and attackers will get tools like it too

Anthropic is limiting who can use this model. But the technique won't stay rare. Security Affairs sums up the stakes: AI capabilities have crossed a threshold that fundamentally changes the urgency required to protect critical infrastructure. Anthropic's own announcement says the window between a vulnerability being discovered and being exploited has collapsed, from months to minutes with AI.

In our view, the window was already shrinking before tools like this arrived. One way to measure it is "time-to-exploit," meaning how long it takes from a flaw being disclosed to attackers using it:

Average time-to-exploit (days), Mandiant
015.7531.547.2563632018-2019322021-2022

From 2018 to 2019, the average was 63 days. Across 2021 and 2022 it dropped to 32 days. In 2023, Mandiant observed an average of just five days.

Now add tools that can turn a published fix into a working attack on their own. Anthropic says the N-day exploits it described were written fully autonomously, starting from just a CVE identifier and a git commit hash. Our read: once a patch is public, assume a capable AI could work backward from it to an attack fast. If your updates wait for the next convenient maintenance window, you're betting that nobody runs that playbook against your equipment first.

Our take: the five-day number is already a historical figure, and AI tools will only push it down. Plan your patching as though the window is days, not weeks.

Your real exposure: edge devices, unpatched SaaS, old plugins

Check which of your tools are software you build yourself and which are products you simply buy and run. But you almost certainly run things that sit on the internet and talk to strangers all day:

Edge devices are the big one. CISA says foreign adversaries routinely exploit software vulnerabilities in network edge devices. Verizon's 2025 DBIR, published April 23, 2025, showed a 34% rise in successful vulnerability exploits, following a 180% rise in the 2024 report. Coverage of the report notes a 34% rise in successful vulnerability exploits, and says exploitation of edge devices and VPNs surged nearly eightfold, from 3% to 22%. Check the latest Verizon DBIR to see how often attackers get in by exploiting vulnerabilities.

Patching is lagging. Of the 17 edge device vulnerabilities featured in the report, Verizon's analysis found only 54% were fully remediated, within a median time frame of 32 days. Compare that to a five-day exploitation window and the problem is obvious.

Old gear makes it worse. Threat actors are increasingly exploiting unsupported edge devices that no longer receive vendor updates. If your firewall is past end-of-support, no patch is ever coming.

This week's example: FortiClient EMS

Here's a real one from this week. The flaw, CVE-2026-35616, is an improper access control vulnerability in FortiClientEMS 7.4.5 through 7.4.6. It may allow an unauthenticated attacker to execute unauthorized code or commands via crafted requests. It earned a critical 9.1 CVSS rating.

Fortinet has observed the flaw exploited in the wild and urges customers to install the hotfix for FortiClient EMS 7.4.5 and 7.4.6. That's three days. The Canadian Centre for Cyber Security advisory recommends that organizations using FortiClientEMS review the Fortinet security bulletin and update or upgrade affected instances.

Heads up

If you, or your IT provider, run this product, ask now whether you're on an affected version and whether the hotfix is in. Don't wait for a monthly maintenance window. This flaw is being exploited in the wild.

A patch-cadence plan for a 10-50 person company

NIST defines enterprise patch management as identifying, prioritizing, acquiring, installing, and verifying the installation of patches, updates, and upgrades throughout an organization. It also frames patching as a cost of doing business. You can't skip it, so the goal is to make it boring and predictable.

Here's the schedule we recommend for a small office. These are our targets, not a standard. Tighten them if you handle health, legal or financial data.

WhatExamplesOur recommended target
Actively exploited flaw on internet-facing gearFirewall, VPN, remote-access, management consolesImmediately, within a few days at most
Critical flaw on internet-facing gear, not yet exploitedSame as aboveWithin a week
Website, CMS and pluginsWordPress core, plugins, themesWeekly, auto-update where safe
Laptops, desktops, browsersWindows, macOS, Chrome, EdgeAuto-update on, verify monthly
Internal servers and line-of-business appsPractice management, accounting, POSMonthly, with a tested rollback
End-of-support devicesOld firewalls, NAS boxes, routersReplace. No patch is coming

Step 1: Build a one-page inventory

You can't patch what you don't know you have. List every device and service that faces the internet: firewall, VPN, Wi-Fi controller, website host, remote desktop tools, cameras, phone system, and any management consoles. For each one, write down the vendor, model, current version, who owns updates (you or your IT provider), and the end-of-support date.

Step 2: Watch the KEV list

CISA maintains the authoritative list of vulnerabilities that have been exploited in the wild, called the Known Exploited Vulnerability (KEV) catalog. CISA says organizations should use it as an input to their vulnerability management prioritization, and recommends a requirement to immediately address KEV vulnerabilities as part of a vulnerability management plan. For a small business, that just means one person (or your IT provider) checks new KEV entries against your inventory every week, and anything that matches jumps the queue.

Step 3: Turn on auto-update wherever it's safe

Check that your browsers, operating systems, office suites, and SaaS apps have automatic updates turned on, and confirm the setting rather than assuming it. Save hands-on patching for the things that can break a workday: firewalls, line-of-business servers, and anything with custom integrations.

Step 4: Put the edge-device SLA in writing

If an IT provider manages your network, ask for a written commitment on how fast they apply critical and actively exploited patches to your edge devices. Ask how they'll tell you when it's done. A slow, fixed patch calendar is no longer good enough for internet-facing gear.

Step 5: Verify, don't assume

After a patch, confirm the version number actually changed and that the old exposure is gone. A scan does this for you.

Tip

Hypothetical: a 25-person accounting firm in Orange County has a firewall, a VPN for remote staff, a WordPress site, and a cloud practice-management app. Under this plan, one person spends about an hour every Monday checking the KEV list against a 15-line inventory, confirms WordPress and plugin updates went through, and calls the IT provider the same day if anything on the edge list matches. That's a small, steady cost that keeps them from being one of the slow patchers.

Using AI-driven scanning on your own systems, safely and with permission

You won't get Mythos Preview. Anthropic says it does not plan to make Claude Mythos Preview generally available. You can still get real benefit without it. The idea: look at your own systems the way an attacker would, regularly, so you find the gaps first.

Start free: CISA Cyber Hygiene scanning

CISA's Cyber Hygiene services are available at no cost. Services typically begin within three business days after sign-up, with reports expected within two weeks of the scanning start date. CISA says U.S.-based federal, state, local, tribal and territorial governments, as well as public and private sector critical infrastructure organizations, are welcome to enroll.

Then: scheduled scanning plus AI triage

A standard vulnerability scanner run regularly against systems you own will catch a lot of what matters for a small office: outdated versions, exposed admin pages, weak settings. In our experience, AI earns its keep one step later. It can read a long scan report, match findings against the KEV list and your inventory, explain each one in plain English, and draft the fix ticket for your IT provider.

If you have custom code, such as a client portal, booking tool or internal app, consider code-level AI scanning too. When you evaluate AI triage tools, look for ones that explain each suggested fix in plain language and let a developer preview, accept, edit, or dismiss it. GitHub says its suggestions were shown to remediate more than two-thirds of found vulnerabilities with little or no editing. That's a vendor benchmark, not a guarantee, but it shows where things are heading.

The permission rule

Our rule: only scan systems you own or have written permission to test. That covers your hosted website (check your host's terms), your cloud apps (check their rules for security testing), and any equipment your IT provider manages under contract. Get permission in writing, set the scope (which addresses, which apps, which time window), and tell your IT provider before the first scan so a sudden burst of traffic doesn't look like an attack.

Pitfalls we see

What we'd do this month

If we were running IT for a 10-50 person office anywhere in Southern California, here's the order we'd go in:

  1. This week: build the edge-device inventory and check it against the KEV list, starting with anything Fortinet, VPN or remote-access related.
  2. Next week: sign up for CISA Cyber Hygiene if you qualify, and set up a scheduled external scan of your public IP addresses and website.
  3. Then: get a written patch SLA from your IT provider, turn on auto-updates everywhere it's safe, and make a replacement plan for anything past end-of-support.
  4. Ongoing: a short weekly check of the exploited-vulnerabilities list, a regular scan, and an AI-assisted summary that turns findings into a short, ranked to-do list.

That last step is where we help. Our pentesting automation and security scanning services (Script Tiger) run permission-scoped scans on a schedule, rank findings by real-world exploitation, and hand you a plain-English fix list instead of a raw report. If you'd rather wire the triage into your existing tools, our AI automation work connects scan results to your ticketing and alerts. You can see how we build security tooling in our Hack House case study.

Next step

AI just showed it can find old, hidden flaws in software everyone trusted. Defenders have access to this capability now, so use it while you can. You don't need a frontier AI model to protect a small business. You need to know what faces the internet, patch it within days instead of weeks, and check your own work.

Want a second pair of eyes on your exposure? Book a Free Automation Audit and we'll walk through your internet-facing systems, your patch cadence, and where scheduled, permission-based scanning would pay off first.

Sources