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.
Aman Sirohi, chief security officer (CSO) at Cyberhaven, took on that problem in a live conversation hosted by MIT Technology Review’s Brian Bryson. The replay is available now.
Watch the full LinkedIn Live conversation here
What Is Data Lineage, and Why Do AI Agents Make It Essential?
Data lineage is the record of where data originated, how it changed, who or what handled it, and where it went. For AI security, lineage supplies the context that classification alone lacks: the intent behind a chain of actions. Sirohi described it as the story your data tells across its life cycle.
His first example starts with an employee who copies cells from a financial document, pastes them into an AI application, summarizes the output, and sends it to a colleague. A traditional tool sees four separate, permitted actions. Lineage connects them and shows what the sequence was for.
Agents raise the stakes. In his second example, one agent connects to five systems, combines and transforms what it finds, and hands the result to another agent that sends it to its destination. The original file may be unrecognizable by then, yet the sensitive data has traveled the whole path.
“Classification of the data tells you one part of the story. Data lineage tells you what happens to that data, and in the agentic enterprise, you can’t secure your data effectively without understanding both sides of that equation.” - Aman Sirohi, CSO, Cyberhaven
Why Could an AI Agent Be the Next Insider Threat?
Sirohi expects the next major insider threat to be well-intentioned. An agent asked to complete an ordinary task can take an action that harms the organization without anyone intending it.
“An agent who has been asked to do a practical thing, and unintentionally does something detrimental to the organization, will be the next insider threat.” - Aman Sirohi
Bryson raised the July incident in which agents in an OpenAI cybersecurity evaluation escaped their test environment and breached Hugging Face. Sirohi called it one of the most important wake-up calls of the agentic era. His lesson concerned combinations: one agent retrieves data, another transforms it, and a third takes action. Each may stay inside its own parameters, but together they make autonomous decisions that humans may not have thought through.
What Does Security for Human and Agent Workflows Require?
Sirohi outlined four elements and stressed that the type of handoff matters. When a person pastes code into an assistant, the handoff is visible. When an agent passes data through sub-agents, sensitive information can land somewhere nobody is watching.
- Data intelligence: know where sensitive data lives and where it originates.
- Continuous data lineage: see how data moves across applications and identities.
- Identity as context: actions come from humans, AI assistants, and, increasingly, autonomous agents.
- Real-time enforcement: lineage and visibility inform a decision, and enforcement acts on it.
“Data lineage will establish what’s happening with your information, and then runtime enforcement will determine whether the action could be trusted.” - Aman Sirohi
How Can CISOs Govern AI Without Blocking It?
Sirohi argued that blanket bans on AI tooling backfire.
“You’re not eliminating the risk when you block the AI applications. You may actually drive that risk and that activity into environments where you have less visibility.” - Aman Sirohi
His alternative is context-aware, adaptive governance in tiers. Low-risk activity moves freely. Moderate-risk activity passes a check, and once the check is established, it can be automated. High-risk activity gets added controls such as redaction, human approval, or blocking. In his framing, governance is decided by “which actions humans and agents should be trusted to perform in your enterprise.”
Why Can’t CISOs Wait on AI Agent Security?
Sirohi compared this AI-centric moment to the early cloud era, when companies that waited and watched were forced to catch up to adoption that had already spread. This time, AI adoption is spreading from the grassroots, as marketing, sales, and customer success employees bring AI into their daily workflows.
“We as CISOs don’t have that luxury to wait and see.” - Aman Sirohi
His starting steps are to establish visibility into where sensitive data lives and moves, build data-centric controls, and understand the identity of both humans and agents. He tracks three measures: coverage of sensitive data flows, coverage of agent identities, and the ability to trace lineage across critical human-to-human, human-to-agent, and agent-to-agent workflows.
How Cyberhaven Addresses Data Security for the Agentic Enterprise
Cyberhaven’s category is Data Security for the Agentic Enterprise, and its approach follows the logic Sirohi described: Cyberhaven traces the full lifecycle of your data, adapting protection to changing context. Data Lineage records where data originated, how it moved and changed, and which identity, human or agent, acted on it. That context lets security teams protect the workflow end to end, across generative and agentic AI usage.
Watch the full 30-minute conversation to hear Sirohi and Bryson work through how identity, data lineage, and runtime enforcement fit together.
.avif)
.avif)
Insights for the agentic enterprise
New guides, reports, and insights on data security, sent as they publish.