Kief Studio

Network Security Monitoring: Detecting AI Browser Anomalies

Protect

  • Published
  • Updated
  • 9 min read
Slow curls of violet and gold smoke rise and fold through a single side beam of light on a black background.

TL;DR: Network security monitoring for AI browser agents works the way it always has: learn what normal looks like, alert on what departs from it, and keep logs the agent cannot touch. What changes is the baseline. Agents make many fast, automated requests, so each one needs its own profile, and hosted agents show up in your app logs more than on your network.

Network security monitoring was built around people and servers. People browse in bursts during work hours. Servers talk to a short list of known places. AI browser agents fit neither pattern. They sign in, fill in forms, call model providers and pull data from many sites at machine speed, sometimes overnight.

That makes the usual network security monitoring alarms noisy in some places and silent in others. This note covers what agent traffic tends to look like, which anomalies are worth an alert, how to build the monitoring with tools most organizations already have, and the legal limits on watching. Everything here applies whether you run agents on your own machines or use one a vendor hosts.

What does an AI browser agent look like on the network?

Glowing AI core inside rings of cyan and violet data, a symbol for an AI agent on the network

Start by asking where the agent runs, because that decides where network security monitoring can see it.

Agents that run on your own machines

A browser agent installed on a workstation or a server in your office produces traffic your network security monitoring can watch. Expect a few things that would look odd for a person: steady connections to the model provider's API, short bursts of many requests while the agent reads pages, automated sign-in flows, and uploads of page content or screenshots for the model to read. None of that is suspicious on its own. It is just different from the human baseline your tools learned.

Agents a vendor hosts

Some agents run entirely in the vendor's cloud. OpenAI's help page for ChatGPT agent says it "will use other tools like its virtual browser to take actions on the web," and that if a task requires a login, it "will pause and prompt you to take control of the virtual browser." (OpenAI Help Center (opens in a new tab)) A browser running on the vendor's side reaches your web apps from the vendor's network, not from your office. Your firewall may never see it. The evidence lands in the sign-in and audit logs of the apps it touches, so those logs belong in your network security monitoring too.

Why does network security monitoring need a baseline for agents?

Sphere of orange gears inside a glowing ring crossed by blue light

Anomaly detection, the heart of network security monitoring, only works against a known normal. CISA's guidance for hardening communications infrastructure puts it simply: "Establish a baseline of normal network behavior and define rules on security appliances to alert on abnormal behavior." (CISA, Enhanced Visibility and Hardening Guidance (opens in a new tab))

NIST's long-standing guide to intrusion detection names "four classes of IDPS: network-based, wireless, network behavior analysis software, and host-based." (NIST SP 800-94 (opens in a new tab)) Network behavior analysis is the class built for this problem: it learns the usual flows and flags what departs from them.

Give every agent its own baseline

Mix agent traffic into the human baseline and you get two bad results. The agent's normal bursts trip alerts all day, and someone tunes the rule down until it misses the real thing. Instead, run each agent under its own identity, on a known host or address where you can, and baseline it separately. Separate identities are also the core of good access control, which we cover in our companion note on identity and access management for AI browser agents.

A per-agent baseline should record which destinations it talks to, at what hours, how much data goes out versus in, and which internal systems it touches. Once you have a few weeks of that, network security monitoring can tell you when a scheduling agent suddenly starts reaching the file server.

Which anomalies point to a compromised agent?

Web of violet lines linking cyan and gold crystal nodes, like hosts mapped by traffic flows

An agent can be turned against you through prompt injection, a malicious extension or a stolen token. We covered the first in prompt injection attacks explained. On the network, the result tends to look like familiar attacker behavior, so the MITRE ATT&CK catalog is a good checklist.

New destinations and beaconing

Watch for connections to destinations the agent has never used, especially on a regular rhythm. Attackers also hide traffic inside ordinary protocols. MITRE notes that adversaries "may communicate using the Domain Name System (DNS) application layer protocol to avoid detection/network filtering by blending in with existing traffic." (MITRE ATT&CK, T1071.004 (opens in a new tab)) An agent host making unusually long or frequent DNS lookups to one domain deserves a look.

Data leaving by an unusual route

Exfiltration is the outcome most owners fear. MITRE describes adversaries who "steal data by exfiltrating it over a different protocol than that of the existing command and control channel," with alternate protocols that include "FTP, SMTP, HTTP/S, DNS, SMB." (MITRE ATT&CK, T1048 (opens in a new tab)) For an agent, the signal is outbound volume that does not match its task: a research agent uploading far more than it downloads, or an agent sending mail when its job never involves mail.

Access outside the task

An agent that suddenly reaches systems outside its job, tries credentials it was never given, or signs in at hours it never works is a strong signal. This is where network security monitoring and access control meet. If each agent is scoped to one task, any step outside that scope is worth an alert.

Resource spikes

Bright orange core at the center of violet and gold structures, like a busy network hub

Sudden jumps in CPU, memory or bandwidth on an agent host can mean the host is doing work nobody asked for. They can also mean a long task or a bug. Treat a spike as a reason to look, not as proof, and check it against what the agent was scheduled to do.

How do you build network security monitoring for agents?

Grid of glowing magenta towers under cyan beams and code panels

You probably own most of the parts of network security monitoring already. The work is pointing them at the right places and feeding them the right logs.

Flow data at the edges

CISA recommends network flow monitoring with exporters and collectors "strategically centered around key ingress and egress locations." (CISA, Enhanced Visibility and Hardening Guidance (opens in a new tab)) Flow records (who talked to whom, when, how much) are the cheapest network security monitoring data to keep, and they are usually enough to spot new destinations and odd volumes without reading content.

Correlation in one place

CISA also advises: "If possible, implement a Security Information and Event Management (SIEM) tool to analyze and correlate logs and alerts from the routers for rapid identification of security incidents." (CISA, Enhanced Visibility and Hardening Guidance (opens in a new tab)) For agents, send four sources into it: network flows, endpoint logs from agent hosts, sign-in logs from your identity provider, and audit logs from the apps agents use. The value is in joining them, so that one event tells you which agent, from which host, acting for which person, did what.

Logs the agent cannot edit

Store network security monitoring data where the agent's own credentials cannot reach it. An agent that can delete its tracks makes every other control weaker. That is a simple rule, and in my experience it is the one most often skipped.

Can machine learning detect AI browser anomalies?

Cyan, gold and magenta waveforms running down a corridor, like traffic over time

It can help, with honest limits. A 2024 peer-reviewed study in the journal AI tested machine learning on network performance data (throughput, latency, jitter, congestion and packet loss). It found that "Unsupervised models effectively captured temporal patterns, while supervised models, particularly Random Forest (94.3%), demonstrated high accuracy in classifying anomalies," and used SHAP values "to enhance model interpretability." (Schummer et al., AI, 2024 (opens in a new tab))

Two cautions. That study measured network performance problems, not attacks, and not AI agents. And a model trained on yesterday's normal will learn an attacker's behavior as normal if nobody reviews what it flags. Machine learning is a good way to rank what a person looks at first. It is not a replacement for that person. Our knowledge base covers threat detection and continuous monitoring in more depth.

What should happen when an alert fires?

Stacked glass cubes holding machinery, wrapped in an orange light trail

Network security monitoring is only useful if someone acts on it. Decide what happens before the first alert, not during it. A short runbook for agents looks like this:

  1. Pause the agent and revoke its tokens. Because it has its own identity, this does not lock out the person it works for.
  2. Preserve the evidence: flow records, endpoint logs, app audit logs and the agent's own task history.
  3. Work out the reach: which systems and data the agent touched during the suspect window.
  4. Fix the cause, whether that is a poisoned page, a bad extension or a leaked token, before you turn it back on.
  5. Decide on notice. If personal information may have been exposed, bring in counsel early.

We walk through the wider version in what happens in the first 24 hours after a breach, and our incident response article covers planning.

Blue electric arc running through dark broken panels, a symbol for an alert

In Massachusetts, monitoring is part of the duty, not just permitted. 201 CMR 17.04 calls for "Reasonable monitoring of systems, for unauthorized use of or access to personal information." (201 CMR 17.04, via Cornell LII (opens in a new tab)) An agent with access to customer records is one of those systems.

Network security monitoring also reaches people, because agents act for staff and process what staff see. Collect only what you need to detect misuse, prefer flow records and logs over full content, set a retention period, and tell employees in writing what is monitored and why. Rules on intercepting communications vary by situation, so have counsel review your monitoring policy before you turn on anything that reads content.

Where should a small team start?

A practical first month of network security monitoring for agents:

  1. List every agent, where it runs and which account it uses.
  2. Move each agent onto its own identity, and onto a known host where it runs on your machines.
  3. Turn on flow logging at your internet edge and on the network segment where agents run.
  4. Send identity provider sign-in logs and the audit logs of your main apps to one place.
  5. Record two to four weeks of baseline per agent, then write alerts for new destinations, unusual outbound volume and access outside each agent's task.
  6. Write the five-step runbook above and test it once with a planned pause.

Frequently asked questions

Can my firewall see what an AI browser agent does?

Only if the agent runs on your network. A hosted agent works in a browser on the vendor's side, so its actions show up in the sign-in and audit logs of the apps it uses instead.

How long should we baseline an agent before alerting?

Two to four weeks of normal work is a reasonable start for most tasks. Rebaseline when the agent's job changes, or its normal will look like an attack.

What is the single most useful alert for agents?

Access outside the agent's task: a system, credential or destination it has never needed. It catches both compromise and misconfiguration.

Do we need a SIEM for network security monitoring?

CISA recommends one where possible, because the value is in correlating network, endpoint, identity and app logs. A small team can start with central logging and add correlation later.

Is machine learning required to detect agent anomalies?

No. Clear baselines and simple rules catch a lot. Machine learning helps rank alerts at volume, but a person still has to review what it flags.

Do we have to tell employees about monitoring?

Written notice is good practice and often expected. Requirements vary by what you collect and how, so have a Massachusetts attorney review your policy.

This article is general information, not legal advice. Kief Studio is not a law firm, and reading it does not create an attorney-client relationship. Laws and regulations change, so check the current text and talk with a licensed Massachusetts attorney about your situation.

Where we fit

Violet waves flowing past glowing cyan bar charts on a dark reflective floor

Kief Studio builds and runs network security monitoring for organizations that have started to put AI agents to work. Our threat detection work sets up flow logging, central log collection and per-agent baselines, and writes alerts a person can act on. Managed security keeps watching after setup, and on what we host, security is included around the clock, as described on LTFI's Watch layer (opens in a new tab). If you want to test whether an agent can be pushed off its task, our penetration testing covers that too.

We work with organizations across Massachusetts, and we will tell you plainly when the tools you already own are enough.

First conversation is free. No commitment.

Tell us which agents you are running

Related articles

  1. Protect

    AI Agent Monitoring: Why Your Security Tools Can't See Your Agents

    Why AI agent monitoring fails: borrowed logins, trusted tools, cloud activity. What to log, how to tell agents from people, and what to alert on.

    Brian Gagne

  2. Protect

    Identity and Access Management for AI Browser Agents

    Identity and access management for AI browser agents: give each agent its own identity, scope it to one task, require approvals and log who acted for whom.

    Brian Gagne

  3. Protect

    AI Audit Trail: What to Record When AI Acts for You

    An AI audit trail proves what an AI did, who approved it and that the record is unchanged. What to record, how to protect it, how long to keep it.

    Brian Gagne