Staying Safe with AI Tools: A Practical Guide

Why This Matters

AI-powered tools — coding assistants, research helpers, browser extensions, and automation apps — are increasingly part of everyday academic work. Unfortunately, the same places we download these tools from have become targets for attackers. Harmful code can hide inside software add-ons, AI model downloads, and browser plugins, and modern AI-assisted attacks are getting harder to spot.

In a perfect world, we'd carefully check every piece of software before installing it, read every line of code an AI writes for us, and review every internet connection our computers make. In practice, that's not realistic — especially in research environments where the pace of work and the need to experiment make strict lockdowns impractical. The good news is that you don't have to do everything perfectly. A handful of easy habits, consistently applied, go a long way toward keeping you safe. This article focuses on those quick wins: being cautious with new software, watching your network traffic, and staying generally aware.

1. Be Thoughtful About What You Install

What's the risk? When you install software — whether it's a Python library, a browser extension, or an AI plugin — you're trusting someone else's code to run on your machine. Attackers take advantage of this by creating harmful software that looks legitimate. AI has made it even easier for bad actors to create convincing fakes at scale.

Yes, we know — IT is always telling you to install updates promptly, and now we're telling you to slow down. The distinction matters: operating system and security updates (from Apple, Microsoft, etc.) should still be installed promptly, because they patch known vulnerabilities that attackers are actively exploiting. The advice below applies to third-party software — libraries, packages, plugins, AI tools, and extensions you download from the broader internet. Those don't go through the same vetting process, and a little patience can save you a lot of trouble.

What to do:

• Wait before grabbing new versions of third-party tools. When a library, plugin, or AI tool you use releases a new version, give it a week or two before updating. Most attacks through compromised software are discovered and taken down within days. Simply waiting would have prevented the majority of major software supply chain attacks in 2025. (This doesn't apply to OS and security patches — install those as soon as they're available.)
• Lock in the versions you trust. If you work with code, specify exact version numbers in your project configuration files (like requirements.txt or package.json) rather than always grabbing the newest version. This stops your tools from silently pulling in a compromised update behind the scenes.
• Be skeptical of unfamiliar software. Before installing something new, take a minute to check: How long has it been around? How many people use it? Is it maintained by a known person or organization? Brand-new software with a generic name and no track record deserves extra caution.
• Stick with well-known tools when you can. When several options do the same thing, lean toward the one with a longer history and a larger community. This is especially true for AI tools, where the landscape is young and fast-moving.

2. Know What Your Computer Is Sending Out

What's the risk? Harmful software often needs to "phone home" — quietly sending your passwords, files, or other information to an attacker's server. If you can see and control what your computer sends over the internet, you can catch and stop this kind of behavior.

What to do:

• Install a local firewall app. Apps like Little Snitch (macOS, paid) or Lulu (macOS, free) will alert you whenever a program on your computer tries to connect to the internet for the first time. Windows users can try GlassWire, and Linux users can use OpenSnitch. These tools put you in control of what gets in and out.
• Watch for unexpected connections. If something you just installed tries to contact an unfamiliar server — especially right after you install or first open it — that's a warning sign. Block the connection and look into it.
• Set it to "ask first." Configure your firewall so that new programs have to ask your permission before going online. By default, most computers allow everything out freely. Flipping this so you have to approve new connections gives you a chance to catch suspicious activity early.
• Check in occasionally. Every few weeks, take five minutes to look through your firewall's log of recent connections. If you see something connecting to a website or server you don't recognize, it's worth a closer look.

3. Stay Aware and Practice Good Habits

What's the risk? Many security incidents happen not because of clever hacking, but because someone ran something without thinking twice. AI has made it easier than ever to create convincing fake emails, fake software, and booby-trapped tools — so a healthy dose of skepticism goes a long way.

What to do:

• Glance at AI-generated code before running it. If you use an AI coding assistant (like Copilot, Claude, or ChatGPT), take a quick look at the code it writes before you run it. Watch for anything that tries to connect to the internet, access files it shouldn't, or looks intentionally scrambled. AI-generated code isn't automatically safe.
• Be careful downloading AI models and plugins. Public download sites for AI models (like Hugging Face or GitHub) and plugin marketplaces have been found to contain harmful entries disguised as useful tools. Check who made it, read reviews, and stick with well-known sources.
• Keep your operating system and security patches up to date. These updates fix known security holes that attackers actively exploit — install them promptly. This is the flip side of the "wait before updating" advice in Section 1: OS and security patches have been vetted by Apple, Microsoft, or your Linux distribution and should not be delayed.
• Try new things in a safe space. When experimenting with a new AI tool or software package, try it in an isolated environment first — a virtual environment, a container, or a separate user account — rather than installing it directly on your main system. That way, if something goes wrong, the damage is contained.
• Trust your instincts about your machine. Unexpected slowdowns, your fan running when you're not doing anything intensive, unfamiliar programs appearing in your task manager, or your internet suddenly being busy for no reason can all be signs that something is wrong. If your machine feels "off," it's worth looking into.
• Report anything suspicious. If you believe your machine may have been compromised — for instance, a tool is making unexpected internet connections or asking for permissions it shouldn't need — contact the Security Operations Center at soc@umd.edu. Early reports help protect the whole community.

Quick Wins — Start Here

You don't need to overhaul your workflow. The items at the top of this list take minutes to set up and cover the most ground. Start with these and build from there as time allows.

None of these practices require deep technical expertise. They're closer to "look both ways before crossing the street" than they are to conducting a full traffic study. Perfect security isn't the goal — making yourself a harder target than the default is, and these steps get you most of the way there.

Further Reading

• OWASP Top 10 for Agentic Applications (2026) — an industry guide to understanding AI-related security risks.
• Little Snitch (macOS) / OpenSnitch (Linux) — firewall apps for monitoring network connections.
• OpenSSF — guidance on evaluating whether open-source software is safe to use.

Questions or concerns about this article? Contact the Engineering Digital Service at eds@umd.edu. To report a suspected security incident or breach, contact the Security Operations Center at soc@umd.edu. This document should be reviewed and updated periodically as the threat landscape evolves.