Cybersecurity in the Age of Generative AI

Generative AI has not rewritten the rules of cybersecurity. It has changed the speed at which they are tested. The organizations that come through this period well will be the ones that keep their fundamentals strong while learning, precisely, where AI genuinely is different.

I have spent more than fifteen years defending financial institutions, technology companies and health systems. In that time I have seen many technologies arrive with promises that everything would change. Usually, the fundamentals held. With generative AI, they still hold — but some of the ground has genuinely moved, and leaders need to know which parts.

01 What stays the same

Most breaches still begin with the basics: a stolen or reused credential, an unpatched system, an account that should have been disabled, a device nobody knew was on the network. None of that changes because an attacker now has a better writing assistant.

• Identity is still the perimeter. Strong authentication, least privilege and tight control of privileged accounts stop more attacks than any new tool.
• You still cannot protect what you cannot see. Complete, current asset inventory remains the foundation of every security program.
• Patching and end-of-life discipline still matter. Old systems with known weaknesses are still the easiest way in.
• People are still the target. Social engineering remains the front door. AI just makes the knock more convincing.

02 What attackers gain

Generative AI lowers the cost and raises the quality of attacks that already existed.

• Phishing without the tells. The spelling mistakes and awkward phrasing we trained people to spot are disappearing. Messages can now be fluent, personalized and produced at scale.
• Voice and video impersonation. A few seconds of public audio can be enough to imitate someone's voice. The urgent call from an executive, a relative or the help desk is now one of the threats I would warn people about first.
• Faster reconnaissance. Gathering and summarizing public information about an organization and its people takes minutes instead of days.
• More capable low-skill attackers. AI assistance narrows the gap between a novice and a practiced operator.

AI has not rewritten the rules of security. It has changed the speed at which they are tested.

03 The genuinely new attack surface

This is where AI security diverges from what came before.

Instructions and data are mixed together

Traditional software separates the code that gives instructions from the data it processes. Large language models do not, by design. Text the model reads — an email, a web page, a document — can carry instructions the model may follow. This class of weakness, prompt injection, has no complete fix today, which means systems have to be designed assuming it will sometimes succeed.

Agents turn words into actions

An AI system that can only answer questions has limited blast radius. One that can send email, move files, query databases or change records is a new kind of user — one that can be manipulated through the content it reads. Every AI agent is a non-human identity, and it deserves the same discipline as a privileged service account: minimal permissions, managed credentials, monitored activity and a human approval step for consequential actions.

Data flows nobody mapped

Prompts, uploaded files, connected drives and conversation histories create new paths for sensitive data to leave the organization or be exposed to people who should not see it. Many of those paths were opened by well-meaning employees, not attackers.

A new supply chain

Models, plugins, extensions and the vendors behind them are now part of your attack surface. Each one deserves the due diligence you would give any third party with access to your data.

Outputs that look right and are not

AI systems can produce confident, plausible and wrong results. When those results feed decisions, code or customer communications without review, the error becomes the organization's.

04 What defenders gain

The same technology is a genuine force multiplier for security teams: triaging alerts, summarizing incidents, explaining unfamiliar code or logs, drafting detection logic for an analyst to review and reducing the hours spent on documentation. Used with review and good judgment, it lets a small team operate like a larger one. Security leaders should be among the most active adopters in the organization, not the most reluctant.

05 Seven moves for security leaders now

• Inventory AI use across the organization, including the tools nobody approved.
• Set data rules for what may and may not go into AI tools, and enforce them where you can.
• Treat agents as privileged identities with least privilege, managed credentials, logging and human approval for high-impact actions.
• Retrain people for the new social engineering, especially voice. Establish verification steps — call-backs on known numbers, code words for urgent requests — that do not depend on spotting a fake.
• Harden the help desk. Identity verification for password and MFA resets is now a frontline control.
• Assess AI vendors on data use, training, retention, access and exit before contracts are signed.
• Join AI projects at the design stage, so security shapes the architecture instead of blocking the launch.

06 The leadership posture

The worst response to generative AI is to treat it as either a miracle or a menace. It is a powerful tool with real risks, and those risks are manageable with discipline. Security leaders who say "no" by default will find AI adopted around them. Security leaders who say "here is how to do this safely" will find themselves at the center of the most important technology decisions their organizations make this decade.