Security | Threat Detection | Cyberattacks | DevSecOps | Compliance

At Cyberhaven, we replaced our previous LLM-based approach to document understanding with a small model that makes 83% fewer decision errors at less than 1% of the inference cost. We achieved this by breaking complex tasks into simpler decisions, using uncertainty as a training signal, and heavily optimizing our inference stack. The model produces calibrated decision probabilities while preserving multimodal understanding and generative capabilities.

An AI support agent picks up a customer case, signs in with valid credentials, works on it in only approved systems, and closes it. I would reasonably predict that if someone were to review this workflow, it would likely come back clean, with no policy rule violation. But if we were to take a closer look, we might find that the agent did more than just close the case.

Security teams built their controls around people, applications, devices, and networks. AI agents operate across all four at once. An agent can pull records from Salesforce, combine them with financial data, summarize the result with a language model, and post it to Slack, with no person touching any step. Each action may look permitted on its own, which means the risk sits in the sequence, and most tools can’t see the sequence.

Data loss prevention (DLP) has been around for decades. But the way data moves today looks very different from the environment traditional DLP was built to protect. That change comes through clearly in the new Market Overview for Data Loss Prevention by Gartner. We see that Gartner describes a mature market undergoing significant evolution, with cybersecurity leaders increasingly looking for DLP that is adaptive, policy-driven, risk-based and AI-enhanced.

An AI agent produces a recommendation that looks reasonable. It's accurate. The model didn't hallucinate. Can the organization trust it? That depends on questions the output can't answer: Enterprise trust requires knowing where information came from, how it moved, what happened to it, and whether the process behind the result deserves confidence. Provenance is where that trust starts.

Every audit season, a governance, risk, and compliance (GRC) team pulls access reports, screenshots configurations, and reconstructs months of data handling from whatever logs survived. The package passes the audit, but the evidence is stale a week later.

Data lineage is fast becoming one of the most important ways to understand data exfiltration, for both human and agentic workflows. More and more vendors now market data lineage, each claiming their approach is superior. That makes us proud parents: what was once an esoteric technology is finally going mainstream. Cyberhaven pioneered data lineage in data security and has spent years perfecting it.

Security teams can spend years tuning DLP rules for the tools they know about. But, then an employee will paste a customer list into an AI assistant that did not exist on any policy six months ago, and the rule never fires. The tool was not on the list, and once the data returns as a summary, it no longer matches the pattern a rule was written for either, creating risk the security team is completely blind to. A bigger rule set does not close this new risk gap.

AI has changed how work gets done. Agents can plan tasks, query enterprise repositories, invoke tools, write code, call application programming interfaces (APIs), and take actions across connected systems. As they work, they create, copy, fragment, transform, and share data across workflows. A perimeter and a one-time authorization decision do not provide enough control for this operating model. An agent can chain permissions, tools, integrations, and data sources while completing a task.

AI didn't replace the old ways data leaves a company. It added new methods on top and gave insiders a way to chain them together. You only see the chain if one sensor watches humans and AI together. If you spend enough time around CISOs / data security/ insider risk practitioners who have run these program for a decade you won't hear that AI is the only thing that matters.

Ask five security leaders what "data security for generative AI" covers and you'll get five different answers. Some may point to an existing DLP rule for ChatGPT, while others would point to an AI acceptable use policy. Some operate under the assumption that their DSPM vendor already handles genAI security. The confusion isn't about effort, as most teams are actively trying to get ahead of GenAI risk.