A few months ago, a security lead I spoke with told me something that stuck with me: “We spent three years building an ISO 27001-certified ISMS, and then someone pasted a client contract into ChatGPT to ‘summarize it faster.'” No malice, no bad intent — just an employee trying to save time. That one moment quietly undid a piece of the control environment the team had spent years building.
This represents the changing role of ISMS today. Generative AI didn’t break in through some exotic exploit. It walked in through the front door, in browser tabs and desktop apps, because it’s useful. And that’s exactly what makes it hard to manage.
Why Generative AI Doesn’t Fit Neatly Into Traditional ISMS Thinking
Most ISMS frameworks — ISO 27001, NIST, SOC 2 — were built around a fairly predictable idea of risk: data moves through known systems, with known owners, known access controls, and known audit trails. Generative AI tools break that model in a few specific ways:
- Data leaves your perimeter voluntarily. Employees aren’t being hacked; they’re choosing to share information with a third-party model to get help with a task.
- The “system” is a black box. You often don’t know how a vendor’s model stores, trains on, or retains what you send it.
- Shadow AI is everywhere. Free AI tools proliferate faster than IT can inventory them, let alone assess them.
- Outputs feel authoritative. People trust AI-generated summaries, code, and policy language without necessarily verifying it, which introduces integrity risk alongside confidentiality risk.
None of this means generative AI is unmanageable. It means the ISMS needs updated eyes.
The Real Risks, Not Just the Headline Ones
It’s easy to jump straight to “employees might leak secrets into ChatGPT.” That may be true,but it doesn’t cover everything.
Confidentiality risk is the obvious one — pasting proprietary code, financial data, or personal information into a public model, where it may be logged, reviewed, or used for training depending on the vendor’s terms.
Integrity risk is subtler. AI-generated content can quietly introduce factual errors, biased outputs, or fabricated information (“hallucinations”) into reports, code, or decision-making processes. If that output feeds into a compliance document or a risk assessment, the error compounds.
Third-party and supply chain risk grows too. Many SaaS tools now embed generative AI features by default, sometimes as an unannounced update. Your vendor risk assessment from last year may no longer reflect what that vendor’s product actually does.
Access and identity risk is often overlooked. AI coding assistants and agents increasingly request permissions to repositories, databases, and internal tools. Overly broad permissions granted to an AI agent can become a new, unmonitored attack surface.
Regulatory and contractual risk rounds it out. Although client contracts, NDAs, and regulations like GDPR may not have anticipated AI-driven data processing, their requirements still remain relevant. Sending regulated data to an AI tool without a data processing agreement can violate obligations you’ve already signed.
Bringing Generative AI Into the ISMS, Not Fighting It From Outside
Banning AI outright rarely works — it just pushes usage underground, which is worse for visibility than sanctioned use would be. A more durable approach treats generative AI as a risk category to be governed, the same way an ISMS already governs cloud services or remote access.
Update the risk assessment, don’t start a new one. Generative AI tools should be evaluated in the same risk register as any other information processing asset — asking what data touches the tool, where it goes, and what could go wrong if it’s exposed.
Classify data before you classify tools. A useful starting question isn’t “is this AI tool approved?” but “is this data allowed to leave our environment at all?” Data classification policies that already exist in a mature ISMS just need to explicitly reference AI processing as a data flow to consider.
Build an AI acceptable use policy that people will actually read. Vague statements like “use AI responsibly” don’t help anyone. Specify what data categories are off-limits, which tools are approved, and what “approved” even means (enterprise agreements with no-training clauses, data residency terms, etc.).
Get visibility before you write more rules. What you cannot see or track is difficult to control effectively. Many organizations are surprised by how many AI tools are already in use once they actually look — through network monitoring, browser extension audits, or simple staff surveys.
Extend vendor due diligence to cover embedded AI features. When reviewing new or existing vendors, ask directly: does this product use generative AI, is customer data used for model training, and can that be disabled contractually?
Train people like the risk is real, because it is. Awareness training that still only covers phishing and password hygiene is incomplete. People need concrete examples of what not to paste into a chatbot, and why.
Treat AI agents as identities, not just tools. If an AI system has API access to internal systems, it needs an access review, logging, and revocation process — the same lifecycle you’d apply to a human employee or service account.
The Mindset Shift That Actually Matters
The organizations handling this well aren’t the ones with the strictest bans — they’re the ones who accepted that generative AI is now part of their information ecosystem, whether it’s officially sanctioned or not. The ISMS doesn’t need to be reinvented; it needs to be extended to a category of tools that move information in ways the original framework didn’t anticipate.
The employee who pasted that contract into ChatGPT wasn’t a conventional security failure. He was a gap in the system — a place where policy, tooling, and awareness hadn’t caught up to how people actually work now. Closing that gap isn’t about mistrust. It’s about giving people safe, sanctioned ways to get the speed and help that AI genuinely offers, so they don’t have to choose between doing their job well and protecting the organization.
That’s the real work of an ISMS in the age of generative AI: not resisting the technology, but making sure sensitive information stays protected no matter which tool touches it next.
