What does cybersecurity look like in the world of AI?
Just because an AI agent running the latest frontier model can build a website doesn't mean you can stop worrying about cybersecurity.
Last month, I attended an information session where the CEO of a cybersecurity software company made a striking point: because so much is being built with AI, there is more risk now than ever. He explained that there are only a few ways to handle risk: transfer it, accept it, avoid it, reduce it, or share it. In today's AI landscape, most risk is being accepted and shared. You build an app, accept the risk that comes with it, and share that risk with providers like AWS or Vercel, who host it on their infrastructure.
All of this new risk is changing the role AI plays in cybersecurity, in three main ways.
First, AI can help modernize old codebases
Much of the internet runs on software that sits idle and is rarely updated. On October 20, 2026, a PHP vulnerability was disclosed that allows an attacker to open a shell on the server. That's significant, because roughly three-quarters of websites run on PHP. Increasingly, vulnerabilities like this are being found and fixed with the help of AI (post on X).
Second, attack methods have changed
As AI spreads, and as people grow more skeptical about AI having access to their personal data, social engineering and phishing are on the rise. It's now difficult to tell AI-generated content from genuine content on social media, and attackers are taking advantage of that. Today's phishing emails look nothing like the obvious scams of the past. Most emails, real or fake, look legitimate, and it takes a careful second look to know which links are safe to click.
Third, the agent has become the target
Attackers no longer go after only servers and databases; they now target AI agents themselves. Personal agents such as Muse and Claude Duet (announced today), agents running inside harnesses like Claude Code, Hermes Agent and Codex, and custom company agents are all connected to a growing number of plug-ins and MCP connectors. These connectors are a newer, more capable kind of API: they don't just give an agent data, they give it the ability to act on your behalf. That makes agent security critical. Many companies are responding by running each agent in its own virtual machine, sandboxing it and limiting what it can reach.
Even a fast System 1 model that only sorts and routes work should get no more access than its job requires.
Looking ahead
AI is changing the world, and every business will need to adapt or be left in the dust. Attackers can now spin up their own agents to probe codebases and cause real damage, which means everyone else has to understand their risks and decide how to handle them. It will be interesting to see what cybersecurity looks like five years from now. Will rogue agents take down entire systems? Will there be a reliable way to tell what's AI-generated and what isn't? It's hard to say. For now, start with the basics: update your PHP, and consider whether it's time to move off PHP altogether.
Adding AI to your business and want it done securely? Start with a free AI Opportunity Assessment.