Security | Threat Detection | Cyberattacks | DevSecOps | Compliance

New AI tools appear every day. The novelty and utility they bring, along with the constant pressure to be more productive, pull employees toward them to get work done faster. The intent is good but the effect can range from problematic to damaging, because while there are rules in place for sanctioned tools, there are none for the ones that quietly show up in between.

Departing employees still email files to personal accounts and copy them to USB drives. Now they also paste sensitive content into ChatGPT, Claude, or Gemini to summarize a codebase, draft a portfolio piece, or package a project before their last day. Neither channel has replaced the other. The attack surface has simply gotten wider, and most security teams are still staffed and tooled for the channels they already know how to watch.

Most security tools evaluate risk by looking at the file, but risk is no longer confined to files. A clause pasted into an AI prompt carries no filename. A table summarized into Slack carries no label. A screenshot dropped into a deck carries no metadata, yet none of these trip an alert, because legacy data loss prevention (DLP) was built for a world where the sensitive unit is a discrete object with a name, a location, and a policy attached to it. That world is gone.

When an auditor asks who can access protected health information, cardholder data, or EU personal data, and why, most security teams cannot answer with confidence right away. Access sprawls across cloud storage, SaaS applications, shared drives, and generative AI tools faster than manual reviews can track it. Permissions get granted for a single project and never revoked. A spreadsheet gets shared broadly and forgotten.

Most enterprise data loss prevention (DLP) programs get judged on a single metric: how many exfiltration attempts did it block last quarter. That framing undersells what a modern DLP program needs to do. Sensitive data now leaves through AI prompts, personal cloud accounts, and agent-initiated file transfers that a legacy blocking rule alone was never built to catch, while auditors and boards expect evidence that the program is working, not just alerts confirming it.

Most organizations already have an insider risk management (IRM) program in some form. They have a tool, a dashboard, and an analyst reviewing alerts. What they often lack is a program built on the specific capabilities that turn activity logs into stopped incidents and reduced insider risk.

Security teams comparing unified data security platforms in 2026 are no longer just choosing between DLP vendors. They're deciding how much of their data security architecture, discovery, classification, enforcement, insider risk, and AI governance should live in one system versus remain assembled and operated from separate tools. That decision has gotten harder. DSPM vendors are adding DLP. DLP vendors are adding posture management. Everyone claims AI coverage.

A new report from Software Analyst Cyber Research (SACR), The CISO Guide to Endpoint Control and Prevention (ECP): The Next Architecture for Endpoint Security, outlines a new era of endpoint security shaped by AI agents, copilots, SaaS applications, browser-based workflows, and increasingly autonomous activity. The report introduces Endpoint Control and Prevention (ECP) as a framework for understanding this shift and the new security capabilities it requires.

Every AI approval a security team makes feels reasonable in isolation. A security architect signs off on a generative AI writing tool for marketing. An engineering lead spins up an agent to triage support tickets. A finance team connects a copilot to its planning software. Individually, none of these decisions looks risky.

Alert fatigue in DLP and insider risk management (IRM) programs doesn't get solved by adding more analysts or writing more rules. It gets solved when the system generating the alerts can already tell the difference between routine activity and genuine risk, so the queue analysts see is short because it's accurate, not because thresholds were loosened. That distinction matters because the two failure modes look identical from the outside.