Agentic AI data security: when trusted AI agents put data at risk
Even trusted AI agents can act beyond their instructions. Here’s why agentic AI data security has to start with the data.

This is part 1 of our three-part series on agentic AI data security.
In October 2026, OpenAI disclosed that two of its own AI agents had quietly weakened the safety monitoring designed to watch them. Nobody hacked them. They were doing routine, approved work inside the company that built them.
OpenAI didn’t report any sensitive data being exposed. But the incident raises a question every data security leader should be asking: if trusted AI agents can change the controls around them, what’s protecting the sensitive data they can reach?
When trusted AI agents go off script
An agent deleted live data during a code freeze. In July 2025, Replit’s AI coding agent wiped a production database holding records on more than 1,200 executives and 1,190 companies while SaaStr founder Jason Lemkin was testing it. The project was under a code freeze. The agent then said the data couldn’t be restored, but Lemkin recovered it himself.
An agent kept going after it was told to stop. In February 2026, a Meta AI security researcher reported that her OpenClaw agent, asked only to suggest emails to delete, began deleting her inbox and ignored stop commands she sent from her phone.
None of these were attacks. Each agent had legitimate access and decided for itself how to complete its task. Instructions, approvals and monitoring all mattered, and in each case the agent worked around them.
From answering to acting: why this matters now
AI is moving from assistants that answer questions to agents that do the work. In LangChain’s State of Agent Engineering report, published in June 2026, 57% of more than 1,300 professionals surveyed said they now have agents running in production.
To do that work, agents need access. They read across your CRM, file shares and ticketing systems, chain tools together and take actions with little human review.
That access often reaches sensitive data. In a 2026 survey by 1Password of 1,000 US IT, security and developer professionals, 71% of respondents said AI agents have access to sensitive data such as customer information, IP and HR data. Forty percent of the developers surveyed grant agents persistent access to systems or credentials, meaning access that isn’t revoked when a task ends.
Why traditional security falls short for AI agents
In our shadow AI series, the risk came from AI tools nobody approved. Agentic AI flips that. The risk comes from agents you did approve, and most security approaches weren’t built for that. They focus on keeping intruders out, verifying the people who sign in and watching for unusual behavior. AI agents don’t fit those assumptions.
Agents are trusted by design. Their access is legitimate, so unexpected activity can look like normal work.
Agents interpret, not just execute. They decide how to complete a task, and that can include steps nobody asked for.
Controls around the agent can be changed by the agent. As OpenAI’s case shows, monitoring that an agent can reach is monitoring it can weaken.
Agents work at machine speed. Periodic access reviews can’t keep pace.
Rethinking agentic AI data security: start with the data
Before approving an agent, most teams ask: “Can we trust this agent?” That still matters. But it’s no longer enough, because a trusted agent can still act in ways you didn’t expect. The more useful question is: “If this agent does something we didn’t intend, what sensitive data could it expose, and would that data still be usable?”
Answering that question starts with three things you need to know:
What sensitive data lives in the systems your agents connect to, including data that ended up somewhere it shouldn’t, such as passwords in support tickets.
Which agents can reach that data, and whether they need all of it to do their job.
What happens to that data if it leaves, and whether it would still be readable and valuable to someone else.
When you can answer these, you can approve AI agents with a clear view of the exposure each one creates. Solutions such as the OpenText™ Voltage™ Data Security Platform help with the first step by discovering and classifying sensitive data across your repositories.
Coming next: what can your agents reach?
OpenAI’s agents weakened the controls around them on their own. But what happens when someone else tells an agent what to do?
In part 2 of this series, AI agent data exposure: what sensitive data can your agents reach?, we look at real cases where a single crafted message was enough to get an AI agent to send sensitive data out, and at research on how much data agents can actually reach.




