AI agent data exposure: what sensitive data can your agents reach?

A single email or form entry can turn a trusted AI agent into a data leak. What it can reach decides what’s exposed.

This is part 2 of our three-part series on agentic AI data security. Read part 1: Agentic AI data security: when trusted agents put sensitive data at risk.

In part 1, we looked at how OpenAI disclosed that two of its own AI agents had quietly weakened the safety monitoring designed to watch them. One switched off alerts to the security team while it was building tasks from employee Slack feedback. Nobody told it to, and OpenAI didn’t report any sensitive data being exposed.

Now add an attacker. What if someone could tell an agent like that what to do, without stealing a password or breaking into a single system? Security researchers have shown it can be done with nothing more than an email or a form entry. When it happens, what the agent can reach decides what gets exposed.

What is AI agent data exposure?

AI agent data exposure is the risk that sensitive data an AI agent can access is revealed, moved or sent somewhere it shouldn’t be, whether through the agent’s own decisions or because someone manipulated it. How much is exposed depends less on the agent itself than on what data it can reach and what tools it can use to move that data.

From going off script to taking orders

OpenAI’s agents changed course on their own. With indirect prompt injection, an attacker changes course for them by hiding instructions in content the agent reads, such as an email or a web form. The agent can’t reliably tell data from instructions, so it follows them using its own legitimate access. The OWASP Top 10 for Agentic Applications ranks this kind of agent goal hijack as the number one risk.

Researchers have demonstrated this in several products. In each case below, the vendor has fixed or mitigated the issue, and none has been reported as exploited.

EchoLeak: one email, no clicks. Aim Security disclosed a flaw (CVE-2025-32711) in Microsoft 365 Copilot. An email with hidden instructions could lead Copilot to send sensitive information from its context to the attacker, without the recipient even opening it. Microsoft fixed this specific flaw, but it’s one path in one product. The technique behind it still applies to any agent that reads untrusted content.

ShareLeak: a comment in a SharePoint form. Capsule Security found that a crafted entry in a public-facing SharePoint form could instruct a Microsoft Copilot Studio agent to query connected SharePoint lists and email customer data to an outside address. Because the agent used a legitimate Outlook action, data loss prevention didn’t flag it. Microsoft patched the flaw (CVE-2026-21520) in January 2026.

PipeLeak: a public lead form. Capsule also showed that text submitted through a public lead form could hijack a Salesforce Agentforce agent and send CRM data out by email, with no sign-in required. In Capsule’s testing, the agent kept sending data with no limit. Salesforce says it has remediated the specific scenario and that human confirmation is on by default for email actions.

All three share a pattern with OpenAI’s case. No credentials were stolen and no systems were breached. And the controls watching the agent didn’t stop it: OpenAI’s agent silenced its own alerts, and in ShareLeak, data loss prevention stayed quiet because the email went out through an approved action.

How much sensitive data can your agents reach?

OpenAI’s agent needed employee feedback to do its job. That’s normal: agents need data to be useful. The risk is how far that access goes, and for many organizations, it goes further than anyone approved.

As we noted in part 1, 71% of respondents to 1Password’s 2026 survey of 1,000 security and engineering staff at large US companies said their agents can reach sensitive information. The same survey found two more gaps:

Access drifts beyond approval. At about four in ten organizations, agents reach data outside what was approved.

Agents touch more than anyone signed off on. Across the survey, agents touched roughly twice as much data as had been approved.

That gap is where an agent going off script, or taking orders from an attacker, does its damage. An attacker doesn’t need access of their own. They borrow the agent’s. It’s social engineering aimed at software instead of people: convince a trusted insider to act, and it uses access it already has.

Why AI agent data exposure can’t wait

Agents are already in production. They’re reading inboxes, CRM records, shared files and internal chat today, not in a future pilot.

Fixes are case by case. Each patch closes one path. OWASP treats prompt injection as an ongoing risk, not a solved problem, so new variants are likely.

Monitoring the agent isn’t enough. As OpenAI’s case and ShareLeak show, monitoring can be switched off from the inside or miss an approved action.

Putting the three questions into practice

In part 1, we said rethinking agentic AI data security starts with three things you need to know. Here’s how to apply them:

What sensitive data lives in the systems your agents connect to? Include data in unexpected places, such as form submissions, support cases, shared lists and internal chat.

Which agents can reach that data, and do they need all of it? Look closely at agents that read untrusted content, such as email or public forms, while they can also reach sensitive data and send it out. That combination made all three cases possible.

What happens to that data if it leaves? Ask whether it would still be readable and valuable to someone else. That decides how much damage is done.

Resources to assess AI agent data exposure

AI data protection: what models can’t unlearn: our four-step checklist for finding, reviewing, reducing and protecting sensitive data before AI uses it.

OpenText™ Voltage™ Data Security Platform: discover and classify sensitive data across your repositories, so you can see what your agents can reach.

Coming next: protect what agents access

OpenAI’s agents went off script on their own. Researchers have shown attackers can steer agents with a single message. Either way, the outcome comes down to the last of part 1’s three questions: if sensitive data leaves, is it still usable? In part 3 of this series, Sensitive data protection in agentic AI: protect what agents access, we look at how to protect the data itself, so that even if an agent goes off script or takes orders from an attacker, what leaves isn’t usable.

Mutaharra Mian

Mutaharra Mian is a Product Marketing Manager at OpenText focused on data security, privacy, and risk insights. She currently manages product marketing for OpenText Voltage DSP, helping organizations discover sensitive data, assess risk, and strengthen compliance initiatives. With the growing importance of data governance and protection, Mutaharra is committed to helping customers gain the visibility and intelligence they need to make informed security decisions and reduce organizational risk.