I spent several years at IBM Security working closely with Guardium. We helped customers discover and classify sensitive data across their databases and data platforms, monitor database activity in real time, and generate the audit trails needed to prove compliance. The model is clear: find the data, label it, watch who touches it, and stop the wrong things from leaving. For structured data sitting in relatively well-defined systems, it worked — and it continues to work well for that purpose.
Then 2023 arrived.
ChatGPT moved from experiment to everyday tool in months. Soon enough, people started pasting meeting notes, customer lists, code, and internal documents into chat windows “just to summarize.” Developers began wiring large language models into business workflows through frameworks like LangChain. Data no longer stayed neatly inside databases or file shares. Prompts, embeddings, third-party APIs, and cloud services had no traditional perimeter. Classic Data Loss Prevention—built mainly for endpoints, email, web gateways, and network traffic—suddenly looked incomplete.
The old playbook assumed you could see and control the channel. Prompt injection, training-data leakage, and unintended model outputs do not play by those rules. An attacker (or a well-meaning employee) does not need to exfiltrate a file. They can steer a model, embed sensitive content in a prompt, or simply ask the right sequence of questions until the model echoes back something it should never have retained or inferred. Public incidents in 2023 and early 2024 made this concrete: system-prompt leaks from Bing Chat, Snapchat’s MyAI, and researchers who coaxed ChatGPT into repeating words until fragments of training data appeared. These were not sophisticated zero-days. They were new paths that traditional DLP was never designed to watch.
IBM’s 2024 Cost of a Data Breach report put the global average at roughly $4.88 million. More important than the dollar figure was the changing nature of the incidents. Organizations faced accidental exposure through generative AI tools, insufficient visibility into how sensitive data was used as context, and growing regulatory pressure. The EU AI Act advanced, CCPA continued to shape expectations in the United States, and other jurisdictions began drafting their own AI rules. Compliance teams could no longer treat data classification as a periodic project. It had to become continuous.
How are you deciding which sensitive data can safely enter a generative AI tool — and which data should never leave a controlled environment at all?
Leave a Reply