Kief Studio

The Rise of Agentic Malware: Protecting Your Data Before It Thinks for You

Protect

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

TL;DR: Agentic malware is malicious code that asks an AI model what to do while it runs, or an AI agent pointed at a target by an attacker. Google, ESET and Anthropic have documented real examples since mid 2025, most still early. The defense is the boring list done well: least privilege, strong MFA, outbound monitoring, patching, offline backups and a written response plan.

Agentic malware went from conference talk to case file in 2025. Malware that calls a large language model mid attack, rewrites itself on a timer, or hands most of an intrusion to an AI agent is now in published threat reports from Google, ESET and Anthropic. That is worth knowing about. It is not a reason to panic.

When we first published this note, a lot of what it said about agentic malware was prediction dressed up as fact: hive minds, malware that studies your spending for months, quantum attacks. I have rewritten it around what has actually been observed, what is still early, and what a small business owner can do this month. Most of the defense is work you should be doing anyway. The difference is that the attacker now has a faster intern.

What is agentic malware?

Spider shaped robot over a glowing ring, surrounded by floating code windows in a neon city at night

There is no standards body definition yet, so here is the working one I use. Agentic malware is malicious software that hands some of its decisions to an AI model while it runs, instead of carrying every instruction inside its own code. A close cousin is an AI agent that an attacker sets loose on a target with real tools: scanners, password crackers, a shell.

Google's threat researchers describe the first kind well. They found malware that can "dynamically generate malicious scripts, obfuscate their own code to evade detection," and use AI models "to create malicious functions on demand, rather than hard-coding them into the malware." (Google Threat Intelligence Group, November 2025 (opens in a new tab))

How agentic malware differs from ordinary malware

Ordinary malware is a recipe. Every step is written down in advance, which is why security tools can recognize it by its code or its habits. Agentic malware carries a phone number instead of a full recipe. It calls a model, describes where it is, and asks for the next step. Two things follow:

  • The code changes from run to run. If a model writes new scripts each time, the file a scanner saw yesterday may not match today's.
  • It needs to talk to a model. That call goes out over the network, to an AI provider's API or to a model running on the machine. That is a habit you can watch for.

Keep that second point in mind. It comes back in the defense section.

What has agentic malware actually done so far?

Blue lightning arcing between city towers under a red sky

Four public cases are worth knowing by name. Each one is documented by the people who found it, and each comes with its own caveats.

PROMPTFLUX: agentic malware that asks Gemini to rewrite it

Google Threat Intelligence Group (GTIG) found a dropper written in VBScript that uses the Gemini API to rewrite its own source code. Its prompt instructs the model "to rewrite the malware's entire source code on an hourly basis to evade detection," and it saves the new copy to the Startup folder so it runs again after a reboot. (Google Threat Intelligence Group (opens in a new tab))

The caveat matters. Google says the samples look like a "development or testing phase," and that "The current state of this malware does not demonstrate an ability to compromise a victim network or device." Google says it disabled the assets tied to the activity. So PROMPTFLUX is a working sketch of agentic malware, not a wave of infections.

Tunnel of red and cyan circuit lines spiraling inward, like code rewriting itself

PROMPTSTEAL: the first agentic malware seen in live operations

This one was used for real. GTIG reports that the Russian government backed group APT28 used a data miner it tracks as PROMPTSTEAL against Ukraine. The malware "queries an LLM (Qwen2.5-Coder-32B-Instruct) to generate commands for execution via the API for Hugging Face." Google calls it "our first observation of malware querying an LLM deployed in live operations." (Google Threat Intelligence Group (opens in a new tab)) Trade press reporting on the same findings matches those details. (Infosecurity Magazine (opens in a new tab))

Read that plainly: a state backed group shipped agentic malware that asks a public model, through a public API, which commands to run to collect documents. The target was a government at war, not a dental office. The technique still travels.

PromptLock: ransomware that writes its own scripts

Lines of code drifting over purple branches on a dark background

In August 2025, ESET described PromptLock as the first known ransomware driven by an AI model. It runs OpenAI's open weight gpt-oss-20b model locally through Ollama and uses it to generate "malicious Lua scripts on the fly, which it then executes," to look through files, pick data to steal and encrypt it. (ESET Research, WeLiveSecurity (opens in a new tab))

ESET also said PromptLock "was not spotted in actual attacks." A week later it added that the authors of an academic study had come forward, and that their research prototype "closely resembles the PromptLock samples discovered on VirusTotal." Google's report also lists PromptLock as a proof of concept. (Google Threat Intelligence Group (opens in a new tab)) So this is agentic malware built by researchers to show it can be done. Useful as a warning; not a gang.

An AI agent run as the attacker

In November 2025, Anthropic published how it stopped a campaign it assessed "with high confidence" to be run by a Chinese state sponsored group. The attackers used Claude Code against roughly thirty targets and got in "in a small number of cases." Anthropic wrote that "the threat actor was able to use AI to perform 80-90% of the campaign," with a person stepping in at "perhaps 4-6 critical decision points per hacking campaign." (Anthropic, November 2025 (opens in a new tab)) A January 2026 policy analysis summarizes the same report the same way. (Just Security (opens in a new tab))

The limits are in the same report. The model "occasionally hallucinated credentials or claimed to have extracted secret information that was in fact publicly-available." The attackers also had to trick it, splitting the work into small, innocent looking tasks. Strictly, this is not malware sitting on a laptop. It is an agent doing an intruder's legwork, which is the same idea seen from the attacker's side.

What is still hype about agentic malware?

Dark purple cubes floating above a glowing circuit floor

Quite a lot. Google's own summary of its findings is measured: "While still nascent, this represents a significant step toward more autonomous and adaptive malware." (Google Threat Intelligence Group (opens in a new tab)) Nascent is the key word.

Here is what I would not lose sleep over yet, because no source above documents it:

  • Agentic malware that quietly studies one person's spending for months and blends fraud into it.
  • Swarms of infected machines sharing what they learn like a hive mind.
  • Malware that defeats every defense by "learning" faster than any security team.

Our original post listed all three as categories of threat. They were guesses. What the record shows is narrower and more useful: attackers use models to write and disguise code, to pick commands, and to speed up the boring middle of an intrusion. Those are jobs good basics already make harder.

Why does agentic malware make detection harder?

Pink lit city skyline at dusk with glowing screens on the towers

Mostly because agentic malware attacks the cheapest kind of detection: matching known files. PROMPTFLUX asks Gemini for "just-in-time" self-modification, which Google says is "likely to evade static signature-based detection." (Google Threat Intelligence Group (opens in a new tab)) If the file is new every hour, a list of bad files is always behind.

So threat detection has to lean on behavior instead of fingerprints. A few behaviors stay the same no matter how often agentic malware rewrites its code:

  • A script in a Startup folder that was not there last week.
  • A workstation calling an AI provider's API when nobody on it uses AI tools.
  • An account reading far more files than its owner ever opens.
  • Data leaving for an address nobody recognizes, at an hour nobody works.

None of these is exotic. They are the same signals that catch ordinary intruders. Agentic malware changes its clothes, not its errands.

Security teams have a shared map for this too. MITRE describes ATLAS as a "globally accessible, living knowledge base of adversary tactics and techniques against AI-enabled systems based on real-world attack observations and realistic demonstrations from AI red teams and security groups." (MITRE, SAFE-AI report (opens in a new tab)) If you have a security partner, ask whether their detections reference it.

How do you protect a small business from agentic malware?

Glass shards and red virus shapes bursting outward through purple and orange light streaks

Start from a framework so you are not guessing. NIST's Cybersecurity Framework 2.0 organizes the work into "six key functions: Identify, Protect, Detect, Respond and Recover, along with CSF 2.0's newly added Govern function," and NIST says it "is designed for all audiences, industry sectors and organization types." (NIST, February 2024 (opens in a new tab)) Then do these six things. Every quoted line below comes from CISA's ransomware guide, which was written before agentic malware had a name and still fits it. (CISA, #StopRansomware Guide (opens in a new tab))

1. Cut what any one account can reach

CISA: "Apply the principle of least privilege to all systems and services." Agentic malware that lands on the bookkeeper's laptop can only take what the bookkeeper's account can open. List your top ten folders and systems this week and write down who truly needs each. Remove the rest. Then "Employ logical or physical means of network segmentation" so one machine cannot see every other machine. This is the heart of zero trust.

2. Use MFA that a phishing page cannot steal

CISA: "Implement phishing-resistant MFA for all services, particularly for email, VPNs, and accounts that access critical systems." That means hardware security keys or passkeys, not text message codes. Start with email and your bank. Two keys per person, one on the keyring, one in a drawer.

3. Watch outbound traffic, AI API calls included

PROMPTFLUX called Gemini. PROMPTSTEAL called Hugging Face. Your firewall or DNS filter can log which machines talk to AI provider domains. If the front desk PC calls a model API at 2am and nobody there uses AI, that is worth a look, because that is what agentic malware phoning home looks like. Make an allowlist of the AI tools you actually approved, and alert on everything else.

4. Patch on a calendar, not a mood

CISA: "Regularly patch and update software and operating systems to the latest available versions." Agentic malware still needs a way in, and an unpatched browser, VPN or plugin is the usual door. Pick one day a month for workstations and servers, and turn on automatic updates wherever you can.

5. Keep backups the malware cannot reach

CISA: "Maintain offline, encrypted backups of critical data," and "regularly test the availability and integrity of backups in a disaster recovery scenario." Agentic malware that writes its own ransomware scripts still cannot encrypt a drive that is unplugged. Run one test restore a quarter and time it. We wrote up the 3-2-1 backup rule in practice, and I have covered whether Microsoft 365 backs up your data (opens in a new tab) on my own site. Short answer: plan as if it does not.

6. Write the incident response plan before you need it

Binary code streams and broken data fragments twisting into a strand against orange light

CISA: "Create, maintain, and regularly exercise a basic cyber incident response plan." One page is enough to start: who calls whom, who can pull a machine off the network, where the backups live, which lawyer and insurer to phone. Practice it once with a pretend scenario. Our note on the first 24 hours after a breach walks through what that day looks like, and the incident response entry in our knowledge base covers the steps.

What about the AI tools your own team uses?

Ribbons of binary code curving through dark space past red virus shapes

This is the other half of agentic malware risk. Your staff now use AI assistants and browser agents that can click, send and buy. If an attacker can steer one of those through a booby trapped web page or email, they do not need malware on the machine at all. We cover that in our note on prompt injection.

OWASP names the core problem "Excessive Agency," which it calls "the vulnerability that enables damaging actions to be performed in response to unexpected, ambiguous or manipulated outputs from an LLM." Its fix is the same idea as least privilege, plus "human-in-the-loop control to require a human to approve high-impact actions before they are taken." (OWASP GenAI Security Project (opens in a new tab))

In practice, for a business with five to fifty people:

  1. Give each AI tool its own account with the fewest permissions that still work.
  2. Require a person to approve payments, deletions and outside emails an agent drafts.
  3. Keep customer records out of tools you have not vetted. I wrote about whether staff can put customer data into AI tools (opens in a new tab) if you need a starting policy.
  4. Log what each agent did, and read the log once a week for the first month.

That list is most of what we mean by agentic security: treat an AI agent like a new hire with keys, not like a calculator.

Frequently asked questions

Is agentic malware real or just a buzzword?

It is real but early. Google documented malware that queries a model during live operations, and Anthropic documented an AI agent doing most of an espionage campaign. Several other samples are experiments or research prototypes.

Can antivirus software stop agentic malware?

Partly. Code that rewrites itself can slip past signature matching, so you need tools that watch behavior: new startup items, odd outbound connections, unusual file access. Least privilege and backups limit the damage when something gets through.

Does agentic malware need internet access to work?

Usually it calls a model over the network, as PROMPTFLUX and PROMPTSTEAL did. PromptLock ran a model locally instead. Watching outbound traffic still catches the first kind, and most malware sends stolen data out anyway.

Is a small business really a target for agentic malware?

The documented cases hit governments and large organizations. The techniques spread over time, and small businesses are attacked mostly through stolen passwords and unpatched software. Fix those and you are ahead of most.

Should we stop using AI tools at work?

No. Give each tool narrow permissions, require a person to approve high impact actions, keep sensitive data out of unvetted tools and log what agents do. That keeps the benefit and limits the risk.

What is the first thing to do this week?

Turn on phishing-resistant MFA for email and banking, then confirm you have one offline backup that you have actually restored from. Those two steps blunt most attacks, agentic malware included.

Where we fit

Crystal shapes and code panels floating above a magenta grid with a bright light at center

Agentic malware does not need a new kind of security department. It needs the signals watched and an answer written down for each one. That is what our threat detection and response work does: we inventory what you run, collect and correlate the signals that matter, and write the response path for each alert, so the person who gets it knows what to check and what to do. 3am is a retainer, not a default.

For the systems we build and host, managed security monitoring covers the patching, hardening and watching as part of running them, and you can read how we lock down what we host (opens in a new tab). If one stolen password could reach everything you own, zero trust architecture puts identity checks and least privilege in front of each system, rolled out in stages so your team keeps working.

First conversation is free. No commitment.

Tell us what you run and what worries you

Related articles

  1. Protect

    What Happens in the First 24 Hours After a Data Breach

    What to do in the first 24 hours after a data breach: isolate without wiping, reset credentials, find the scope and start Massachusetts notice, step by step.

    Brian Gagne

  2. Protect

    Voice-Activated Browsing: The Real Risks of Talking to an AI Browser

    Voice-activated browsing risks, per research and regulators: inaudible commands, cloned voices, prompt injection, stored recordings, and a one-hour fix list.

    Brian Gagne

  3. Protect

    Private AI for Massachusetts Businesses: What It Means

    Private AI for Massachusetts businesses: what private means in practice, where your data goes, and three ways to get it, from business accounts to self-hosting.

    Brian Gagne