Your Business Is Already in an AI Cybersecurity War. Here's How to Not Lose It.

Last week, Anthropic revealed that an experimental AI model called Mythos had found exploitable zero-day vulnerabilities in every major operating system and every major web browser. The same week, the FBI warned that Iranian-affiliated attackers were actively disrupting US water and energy facilities, and reported that cybercrime losses hit a record $20.87 billion in 2025. If you run a business, none of that is background noise. The window to harden your basics before AI-scale attacks become routine is closing, and it was already open longer than it should have been.

Here's what the last seven days actually mean for your operations, and what I'd do about it this week if I were in your shoes.

The attackers just got a capability preview

Let me be precise about what happened, because the framing matters. Anthropic's Mythos model is not loose in the wild. The company deliberately held it back and distributed a preview version only to a small group of large vendors through something called Project Glasswing. Anthropic's own assessment is that releasing it publicly would "break the internet."

What we do know, from Anthropic's disclosure as reported by The Register and Wired, is that in internal testing Mythos generated working exploits 72.4% of the time, compared to a near-zero rate for the current Claude Opus model. Researchers documented it independently building a four-vulnerability browser exploit chain, a Linux privilege escalation using race conditions, and a FreeBSD remote code execution through a 20-gadget ROP chain.

That is the kind of work that used to take an elite security team weeks or months. An AI model produced it in a session. The full production version is held back, but the capability has been demonstrated publicly. Every well-resourced attacker on the planet is now trying to reproduce it, and the gap between "lab demo" and "commodity tool" in AI has historically been measured in months.

Meanwhile, the real world didn't wait for a lab. The FBI and CISA issued a joint alert that Iranian-affiliated actors have been targeting Rockwell Automation and Allen-Bradley programmable logic controllers across water treatment plants, oil refineries, power grids, and food production sites. Some victims have already experienced operational disruption and financial loss. These are not AI-powered attacks. They are the same attacks we've seen since 2021, exploiting the same default credentials on the same internet-exposed PLCs. The ports those attackers probe (44818, 2222, 102, 502) sit in more small and mid-sized industrial networks than most owners realize.

That's the uncomfortable part. The attackers don't need Mythos yet. The old attacks still work, because most businesses never closed the 2021 gaps. Now imagine those same attackers with AI-generated exploit chains on top.

The defenders are also AI. You're probably not on the invite list.

The same Mythos model is being used defensively. Anthropic paired with Nvidia, Google, AWS, Apple, Microsoft, Cisco, and over forty other organizations to stress-test their own products and patch vulnerabilities before bad actors get there. The company committed up to $100 million in usage credits and $4 million in donations to open-source security groups to support the effort.

Notice who's on that list. The hyperscalers, the operating system vendors, and the companies whose names end up on the front page when a breach happens. Your small-to-mid sized business is not in that room. You're downstream of whatever those vendors ship, and you'll get the patches when they arrive, which is usually later than you want.

That's not a complaint. It's the shape of the landscape. And it tells you exactly where your attention should go this quarter.

The numbers already reflect the shift

The FBI's 2025 Internet Crime Report crossed the $20 billion line for the first time ever, with losses of $20.87 billion and over a million complaints, up 17% year over year. Inside those numbers, a few lines stand out for anyone running a company.

Government impersonation scams jumped 128% in two years, from around 14,000 reports to over 32,000. These are the deepfake voice calls and messages pretending to be the IRS, a federal agent, or a local official, and they are landing on employees, customers, and vendors. Interpol researchers cited in the report found that fraud schemes using AI are 4.5 times more profitable than those without. That's not a projection. It's already in the ledger.

Investment scams alone accounted for $8.6 billion in 2025 losses. Business email compromise and tech support fraud rounded out the biggest categories, and cyber-enabled fraud now represents 85% of total financial losses while only making up 45% of complaints. The scams that hurt the most are the ones that use your own systems and relationships against you.

On top of that, the White House is proposing a $700 million cut to CISA, the federal agency most businesses quietly depend on for threat intelligence and incident coordination. Whatever you think of the politics, the practical effect is the same. Less federal air cover. More responsibility on you.

What to actually do this week

None of the advice below is new. MFA, incident response plans, staff training. This is the same checklist good security consultants have been writing since 2018. That's exactly the point. The fundamentals haven't changed. The urgency has. Here's the shortlist I'd run through if I were running any small or mid-sized operation right now, in order of impact.

Audit your multi-factor authentication, for real. MFA on email is the minimum, not the finish line. Make sure it's on every admin account, every remote access portal, every payroll and accounting tool, and every vendor portal you use. If any of those still accept a password alone, that's where the next breach starts.

Ask your IT provider or MSP one specific question. "What's your response plan for an AI-assisted attack on our environment, and can you show me where it's documented?" If the answer is hand-waving, that is your answer. You either get a real plan back, or you start looking for a provider who has one.

If you run anything industrial, segment and harden your PLCs. I know you can't always take them fully off the public internet. Vendor remote maintenance exists. SCADA interoperability exists. Fine. Then put them behind a jump host with MFA, segment the OT network from the rest of the business, rotate every default credential, and check logs for suspicious traffic on ports 44818, 2222, 102, and 502, especially from hosting providers overseas.

Train the humans on deepfake voice and impersonation scams. A twenty-minute team meeting this week is cheap. The template is simple. No financial action, no credential sharing, and no vendor change gets processed off a voice call or an email alone without a callback on a known number. Write it down and stick it on the wall.

Review your incident response plan and time-box it. When something happens, who gets called, in what order, within how many minutes. If you don't have this written down, write it today. One page is enough. Then actually run a tabletop exercise with your team before the end of the month.

If you don't have cyber insurance, price it this week. Premiums are moving, and most providers are tightening their underwriting questions. The sooner you engage, the cleaner the terms you're likely to get.

The window is still open. It is not going to stay that way.

The story I keep coming back to from this week is not the Anthropic research paper, impressive as it is. It's this: Anthropic showed a preview of a model that finds zero-days at a 72% success rate, and in the same seven-day window, Iran was still owning PLCs with default passwords, and the US was still logging over $20 billion in cybercrime. The old gaps are wide open, and the new capability is visibly on the way.

The businesses that treat the next month as a real window to close the basics are going to be fine. The businesses that assume their current stack will absorb the shift, because it always has, are going to discover something uncomfortable. The attackers don't need Mythos to hurt you today. Closing the 2021 gaps is the only thing that gives you a chance once the 2026 ones arrive.

Pick one item from the list above and do it today. Not next quarter. Today.