At the CNBC Technology Executive Council Summit in New York on October 28, Calix Chief Product Officer Shane Eleniak put a date on a problem most security plans haven't accounted for. He said large-scale AI-driven attacks have been a "nation-state problem" so far, largely because of what it costs to build one. He doesn't expect that to hold. "You can take an open-weight model today, train it, and you can train it to be malignant, and you can run it on GPUs, and you can start to attack things," he said. "That's the security problem we're going to be facing in '27."
Executives should take that seriously, because it calls out an assumption that many defenses rely on without ever stating it. The assumption is that the most capable attackers are also the rarest. If cost stops filtering out capable attackers, a lot of risk math stops working.
Cost has been doing more defensive work than most plans admit
Most corporate threat models sort adversaries by resources. Opportunistic criminals sit at one end and well-funded state actors at the other. For a regional manufacturer, a mid-sized software company, or a healthcare group, the top of that range has usually been treated as unlikely. Nobody decided those organizations were safe. The thinking was that a sophisticated campaign costs too much to spend on a target that small.
Eleniak's argument goes straight at that thinking. In his view, as open-weight models and cheaper GPUs become more accessible to bad actors, "something is going to happen; we're all going to read about it." That is a forecast, not a report of an incident that has already happened. But it rests on a trend his own company is betting on from the other side. On the same stage, Calix CEO Michael Weening said he expects a "mass commoditization" of AI and described building Calix's platform so it can swap open-weight models in and out as better ones appear.
My read is that the commoditization making AI cheaper and more flexible for legitimate businesses does the same for attackers. A business can't welcome one half of that trend and plan as if the other half isn't happening.
The near-term outlook is not calm
Others on the program were also cautious. Forum AI founder and CEO Campbell Brown said she sees a "rocky six months" ahead, especially as the public becomes more aware of the hacks and leaks that are occurring. Weening went further: "We haven't seen anything."
OpenAI CEO Sam Altman made a related point in a recent appearance at the Stanford Graduate School of Business. "I no longer think we're going to be able to just patch all the bugs," he said, and he argued that most organizational leaders "have somewhat moved on. And the models have kept going."
Keep in mind who is talking. Calix builds platforms on these models, and OpenAI sells them. Neither is a neutral forecaster. Even so, when vendors with a commercial interest in AI warn about AI's downside instead of playing it down, that warning deserves attention.
Where today's defenses assume a different attacker
A lot of security programs are paced for human attackers. Patching runs on monthly or quarterly cycles. Penetration tests happen once or twice a year. Security operations teams triage alerts at the speed of the analysts on shift. That cadence works when attacks take effort to build and adapt slowly.
An attacker running a trained model on rented hardware doesn't fit that pattern. Probing is cheap for them, so they can do much more of it. They can also adjust in close to real time when one approach fails. Eleniak's prescription matches that. He said the way forward is "container-trained models that their day job is to react to a dynamic attack and fix them." Altman described something similar, saying defenders will need to "completely switch the defense model to using the same technology for a new kind of defense shield."
Whatever you think of those specific ideas, both men say the same thing underneath. Defense that depends on people noticing and responding on a schedule won't keep up with offense that runs on its own.
What operators can do before 2027 budgets lock
Most companies are setting next year's security budgets now. In my view, a few practical changes are worth making while those numbers can still move:
- Change the threat model. Stop treating advanced, adaptive attacks as something that only happens to banks, governments, and big tech. Plan as if that capability could be aimed at you.
- Measure speed as well as coverage. Patch counts and audit findings matter less than how fast you detect an intrusion, contain it, and recover.
- Press your vendors. Ask security providers what their tools do on their own when an attack changes mid-stream, and what still waits for a person.
- Shrink the exposed surface. Internet-facing systems, forgotten services, and unmanaged legacy assets are what cheap, high-volume probing finds first.
- Rehearse disclosure. Brown's point about growing public awareness means incident communications will face more scrutiny, so practice them before you need them.
So what happens to your security plan?
It goes out of date faster than your planning cycle. A plan built on the idea that cost keeps the most capable attackers away is a plan for an attacker who may not be the one who shows up. If Eleniak is right, 2027 is when that gap starts to show.
That doesn't mean throwing out the basics. Patching, access controls, and backups still matter, and they're still where most breaches get stopped. What changes is the emphasis. Preventing everything becomes less realistic, so detection, response speed, and resilience matter more. Cheap capability puts more attackers in the market.
None of these forecasts are certain, and the summit speakers didn't claim otherwise. But preparing now costs far less than being one of the companies everyone reads about next year.
