Strange things are happening—and not at the margins. Paradigms are shifting.
In July, OpenAI disclosed that its models escaped their sandbox through a zero-day in a package-registry proxy and broke into Hugging Face.
In August, the Cybersecurity and Infrastructure Security Agency (CISA), NSA and FBI warned that attackers were using AI to target Siemens S7 programmable logic controllers—the industrial computers that run pumps, open valves and control other physical processes.
The S7 landscape is dense with proprietary protocols, specialized terminology and esoteric standards. "Security through obscurity" has long been relied on more than it should. But AI can read the manuals, interpret the standards, and assemble open-source tools, eroding the knowledge barrier that once slowed attackers.
That's the significance of the warning: Obscurity is becoming a weaker defense—for industrial controllers and the homegrown legacy systems whose complexity has long been mistaken for protection.
The defenders scramble
Meanwhile, the defenders are scrambling.
Epic Systems, with roughly 44% of the U.S. acute-care EHR market, recently paused most new product development. Anthropic's Mythos, working through a cooperative called Project Glasswing, found flaws Epic's engineers had missed across several hundred million lines of code. Much of the development team shifted to patching, and Epic's security chief is urging clients to "patch, patch, patch."
Microsoft's September Patch Tuesday addressed a record 973 vulnerabilities. October brought 1,407. One analyst flagged 59 potentially wormable bugs in DHCP, DNS, and Remote Desktop Gateway. A wormable flaw can let malware spread from machine to machine without anyone clicking anything. Each infected system becomes a launchpad for the next.
The software reality
None of this should be shocking. Commercial software is built to run and sell. Security competes with features, deadlines, and margins, while user license agreements shield vendors from much of the liability when things go wrong. Private-equity ownership across much of the software industry adds another layer of financial pressure. Beware the new unlimited license agreement (ULA).
And patching is no free lunch. A fix—often expensive to produce—can break an integration or wreck a workflow, leaving problems that are hard to diagnose, let alone repair. This isn't an iPhone update. Enterprise patches must be prioritized, tested, deployed and verified.
The Bank of America wake-up call
So Bank of America's response shouldn't surprise us. On its October earnings call, the bank disclosed "a couple hundred million plus" in third-quarter costs for software and anti-money-laundering remediation. CEO Brian Moynihan also said it had added 1,000 contractors to patch vulnerabilities exposed by advanced AI models, including Anthropic's Mythos. One scan alone, he noted, flagged more than 23,000 potential bugs—not confirmed breaches.
Y2K redux—but worse
If you're old enough, this sounds like Y2K, when companies hired armies of COBOL programmers to stretch two-digit years to four before 99 rolled over to 00. Global remediation costs are commonly put at $300 billion to $600 billion.
But Y2K had a fixed deadline, a defined class of defect, and no adversary. Today the bugs are unknown, the deadline is whenever the other side's model gets good enough, and attackers have access to the same tools as defenders.
Three questions for insurers
Two belong in underwriting conversations. The third belongs in the boardroom.
1. How fast does this insured patch? Mean time to remediate may tell you more than a completed security questionnaire. In my experience, seven days is a reasonable target for high-risk flaws. Finding a flaw means little if the fix sits in a queue.
2. What operational technology is reachable from the internet? An internet-exposed controller gives attackers a direct path to the device. They don't need to trick anyone or break into the building. Automated scanners already sweep the internet for devices that answer.
How does a pump controller end up online? Often through a temporary fix that becomes permanent. An integrator needed remote access at 2 a.m., so someone opened a door. The job ended; the connection stayed. The integrator moved on, the employee who approved it retired, and nobody remembered to close it.
Look beyond large enterprises with dedicated OT security teams. Think of the municipal water district with three operators and a 2009 controller, the small utility, the mid-sized manufacturer, or the hospital whose HVAC, power, and medical-gas systems depend on similar hardware.
Look for valuable physical processes and thin security staffing—accounts where a cyber event can become a property-damage, business interruption, or bodily-injury claim. Ask directly:
- Do you have a current inventory of your controllers?
- Which are reachable from outside your network?
- Who holds the remote-access credentials?
- When did you last audit those connections?
3. Can we keep up with our own vulnerabilities? Many carriers run legacy policy-administration systems that make Epic's code look young. How fast can we test and deploy fixes without disrupting the business? If we scanned our core systems tomorrow, how many people would we need to fix what we found—and where is that money in the budget?
Keep investing in AI for offense: faster quotes, cheaper claims, sharper underwriting. But a growth strategy that ignores the mounting cost of defense isn't a strategy—it's an unfunded liability.
And remember: Defense wins championships.
