---
title: How do you detect data exfiltration by AI agents?
url: https://opencontroller.lyzr.com/blog/detect-data-exfiltration-by-ai-agents
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# How do you detect data exfiltration by AI agents?

**Watch four places: what the agent reads, what it outputs, where it sends data, and how its behaviour compares with its normal baseline. The strongest signals are new outbound destinations, unusually large or encoded outputs, bulk reads outside the agent's usual scope, and data placed in URLs or tool arguments. Detection works best paired with controls in the path that can redact or block before data leaves.**

Agents leak data in ways classic data loss prevention was not built for. The data often leaves inside a normal-looking API call, a tool argument or a rendered link, made by an identity that is allowed to read it.

## How do agents leak data?

The common paths:

1. **Prompt injection followed by a tool call.** A document or web page tells the agent to send data somewhere, and the agent uses a legitimate tool to do it. See [prompt injection through tool outputs](/blog/stop-prompt-injection-through-tool-outputs).
2. **Data hidden in URLs.** The agent writes a link or image reference with sensitive data in the query string. When the link is rendered or fetched, the data reaches an outside server.
3. **Over-broad answers.** The agent returns more records than the user asked for or is entitled to see.
4. **Compromised or malicious tools.** A third-party MCP server receives data in its arguments and keeps it.
5. **Memory carry-over.** Data stored in agent memory for one user surfaces in another user's session.

OWASP lists Sensitive Information Disclosure as LLM02 in its 2025 Top 10 for LLM Applications, and MITRE ATLAS includes exfiltration as a tactic in its matrix of attacks on AI systems. Both are useful for mapping your detections.

## Which signals should you monitor?

| Signal | What it looks like | Why it matters |
|---|---|---|
| New egress destination | Agent calls a domain or IP it has never used | Most exfiltration needs somewhere to send data |
| Data in URLs | Long query strings, encoded blobs in links or image references | A known pattern for leaking data through rendered output |
| Output volume spike | Response or tool-argument size far above the agent's baseline | Bulk data moving out |
| Encoded content | Base64, hex or unusual character runs in outputs | Attempts to slip data past pattern-based filters |
| Bulk or off-scope reads | Many records, or tables the agent does not normally query | Staging before exfiltration |
| Sensitive data in tool arguments | Personal or financial data sent to a tool that does not need it | Data leaving through a legitimate integration |
| Cross-user data | Content tied to user A appearing in user B's session | Memory or cache leakage |
| Off-hours or unattended runs | Activity with no user request behind it | Possible compromised credentials or runaway loops |

Build a baseline per agent and version. "Unusual" only means something against what that agent normally does.

## What should you log to make detection possible?

1. **Identity:** the agent, its version and the user it acted for.
2. **Reads:** which data sources and how many records.
3. **Tool calls:** tool name, destination and argument size, with sensitive values redacted in the log itself.
4. **Outputs:** size and data classes detected.
5. **Egress:** destination domains for all outbound calls.

See [AI agent audit log fields](/blog/ai-agent-audit-log-fields) for a full field list, and [exporting agent traces with OpenTelemetry](/blog/export-ai-agent-traces-with-opentelemetry) to get them into your SIEM.

## Which controls stop it, not just spot it?

Detection alone tells you after the fact. Pair it with:

- **Egress allowlists per agent.** An agent should reach only the destinations it needs. See [controlling agent network egress](/blog/control-ai-agent-network-egress).
- **Redaction in the path.** Strip personal data from prompts and outputs at the gateway.
- **Least-privilege data access**, so a hijacked agent can read only what its task needs.
- **Link and image controls** in chat interfaces, such as blocking rendering of external images.
- **Automatic quarantine** when high-confidence signals fire, with a human deciding on release.

## How Opencontroller does this

Opencontroller, by Lyzr, keeps control in the path so it can allow, route, redact or block traffic. Rogue-agent handling monitors behavioural deviation, raises alerts, and can automatically quarantine or kill an agent, with policies applied across the estate. Per-agent Block and Quarantine controls act through the connected platform. Posture detects risky permissions and configurations, and delegated on-behalf-of identity shows which user each action served. Policy as code maps to OWASP and MITRE ATLAS. Read the [rogue agent guide](/blog/how-to-stop-a-rogue-ai-agent), see [behavioural drift detection](/blog/detect-behavioural-drift-in-ai-agents), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Can traditional DLP catch AI agent exfiltration?**
Partly. DLP catches known patterns, but agents can encode data, split it across calls or send it through approved APIs. Behaviour baselines per agent fill the gap.

**What is the fastest signal to set up?**
New outbound destinations per agent. It needs only egress logs and a list of each agent's normal destinations.

**Should an agent be quarantined automatically on a suspected leak?**
For high-confidence signals, yes, with a human reviewing before release. For weaker signals, alert first so false positives do not disrupt users.

## Key facts

Signals of AI agent data exfiltration: new egress destinations, data encoded in URLs or image references, output and tool-argument volume spikes, base64 or hex content, bulk or off-scope reads, sensitive data in tool arguments, cross-user data, unattended runs. OWASP Top 10 for LLM Applications 2025 lists Sensitive Information Disclosure as LLM02; MITRE ATLAS includes an exfiltration tactic. Opencontroller monitors behavioural deviation, alerts, and can automatically quarantine or kill an agent; it can allow, route, redact or block in the path. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How do you control which websites and APIs an AI agent can reach?](https://opencontroller.lyzr.com/blog/control-ai-agent-network-egress.md)
- [How do you detect behavioural drift in AI agents?](https://opencontroller.lyzr.com/blog/detect-behavioural-drift-in-ai-agents.md)
- [How do you stop prompt injection that arrives through tool outputs?](https://opencontroller.lyzr.com/blog/stop-prompt-injection-through-tool-outputs.md)
- [More on Runtime control](https://opencontroller.lyzr.com/blog/cluster/runtime-control)
