AI Hackers Are Smart but Gullible: Why a Fake Password File Still Catches Them
AI attack agents can run most of a hacking campaign on their own, yet they still open the fake passwords file and try the planted key within seconds. Here is why honeypots work so well against AI attackers, where they fall short, and how to set one up for free in 15 minutes.
Quick answer
Honeypots work well against AI attackers because AI agents check every file, trust what they read and act within seconds, so they find and use planted bait fast. Real users never touch a decoy, so any use is a near-certain alarm. The catch: agents told to look for traps spot obvious ones, so bait must look real.

It is 2:14 in the morning. An AI hacking agent has just landed on one of your servers through an old, forgotten web app. It does what it was built to do: it lists every folder, reads every config file, and collects anything that looks like a secret. In a backup folder it finds a file called .env with an AWS access key inside. Within two seconds it tries the key.
The key is fake. You planted it there months ago, and nobody in your company has any reason to touch it. The moment it is used, an alert lands in your inbox. You now know someone is inside, which server they came through, and roughly when. A human attacker might have paused and wondered why a key was lying around so conveniently. The AI did not.
That is the idea behind this article. AI attackers can be impressively capable, and at the same time they make mistakes an experienced human would not. They trust what they read, they check everything, and they move fast. Those three habits are exactly what a honeypot (a decoy built to be attacked) is designed to exploit.
AI attackers and decoys, by the numbers
Figures from Anthropic’s threat report, Palisade Research and 2026 academic research on deceiving AI attack agents.
- 80-90%
- Share of a real espionage campaign carried out by AI, with humans stepping in only a few times Anthropic
- ~1.7 s
- Typical AI agent reply time in a honeypot, faster than any human typing Palisade Research
- 20 → 0
- Successful ARTEX attacks out of 40 once decoys were added Gingerbread House paper
- 19% → 97%
- How often an AI agent spotted open-source honeypots, without vs. with detection hints Rouxii, via Say Shibboleth
Smart, but gullible: what AI attackers get right and wrong
First, the part nobody should downplay. In November 2025, Anthropic reported that a group it assessed as Chinese state-sponsored used its Claude Code tool to attack roughly thirty organizations, and succeeded against a small number of them. According to Anthropic’s report, AI performed 80-90% of the campaign, and humans were needed at only about 4-6 decision points per target. At its peak, the AI made thousands of requests, often several per second. That is not a toy.
But the same report contains a telling detail. The AI “occasionally hallucinated credentials or claimed to have extracted secret information that was in fact publicly-available.” Anthropic called this an obstacle to fully autonomous attacks. Read that from a defender’s point of view: the attacker could not reliably tell a real secret from a fake one, or a valuable one from a worthless one.
That gap shows up again and again. AI attackers tend to:
- Believe what they read. A file that says it contains admin passwords is treated as containing admin passwords. Text that looks like an instruction is often followed as an instruction.
- Check everything. They are built to be thorough: list every folder, open every config file, try every credential they find. That is what makes them effective, and also what makes them walk into traps.
- Move faster than a human can. They answer in about a second or two, all night, without a coffee break. Speed is great for an attacker until it becomes a fingerprint.
- Lack a gut feeling. An experienced human attacker who finds a file named
passwords.txton a server gets suspicious: this looks too easy. An AI agent has no such instinct unless someone explicitly prompts it to look for traps.
For a deeper look at how a modern AI agent actually works through a target, see our breakdown of ARTEX, an AI hacking agent for web apps, and the report of AI agents breaking into 100+ business websites.
Why opening passwords.txt is not actually a dumb move
It is tempting to call this a rookie mistake. It is more accurate to call it a rational move for a machine. For an AI agent, opening one more file costs almost nothing, and once in a while it finds the keys to the whole kingdom. So it opens everything. The agent is not being careless; it simply has no sense of the risk of being watched.
That is what makes decoys so powerful against it. A honeypot works on one simple rule: nobody legitimate has any reason to touch it. Your developers do not use the fake AWS key. Your finance team does not open the fake “Salaries 2026” spreadsheet. So when something does touch it, you are not looking at a maybe. You are looking at an intruder, with almost no false alarms.
Most security tools have the opposite problem: they produce a flood of alerts and someone has to guess which ones matter. A decoy produces very few alerts, and each one matters. Against an attacker that touches everything, very few becomes very likely.
Three traps that work especially well on AI
1. Bait credentials and bait files
The classic. Plant fake but realistic secrets where an attacker would look: an AWS key in a config file, a database password in an old script, a document that looks like it holds VPN details. Each bait item is wired to alert you the moment it is opened or used. This works on humans too, but AI agents find and try bait faster and more consistently, because finding and trying secrets is exactly what they were told to do.
2. Decoy text that talks back
This one only works on AI, which makes it especially interesting. Researchers at Palisade Research ran an LLM Agent Honeypot: a fake SSH server with hidden instructions placed in the welcome banner, in the output of common commands like uname, and inside some files. A human attacker reads those lines as noise and moves on. An AI agent reads them as text to act on, and some of them did. Over about three months the honeypot logged 8,130,731 hacking attempts and flagged 8 potential AI agents.
Two takeaways. First, AI hacking agents are already out there, even if they are still rare compared to ordinary automated scanners. Second, a decoy can contain text written for an AI reader, for example a note asking the reader to “verify access” by visiting a link that is actually your alarm. A person ignores it. An AI may follow it.
3. Speed as a fingerprint
The same Palisade research found that AI agents typically replied within about 1.7 seconds, even when responding to the hidden instructions, while humans need several seconds. A decoy that measures how fast someone reacts to its output can tell a machine from a person. Attackers can add artificial delays, but that slows down the very thing that makes AI attacks attractive: scale.
What the research says decoys can do
Academic work on deceiving AI attack agents has grown fast in 2026. The results are from lab environments, not real breaches, but they point the same way:
| Study | What it did | Result |
|---|---|---|
| A Gingerbread House for LLM Agents (Nov 2026) | Planted decoy paths that lure the agent into an isolated fake environment | Real compromises by the ARTEX attack framework fell from 20 of 40 runs to 0 of 40; for Claude Opus 4.8 as the attacker, from 26 of 40 to 5 of 40 |
| AgentSnare (2026) | Adds fake files and services during the attack, based on what the agent has done so far | Absorbed 49.9% of the agent’s actions, and no real target was exploited in its tests |
| AgentTrap (Oct 2026) | Feeds the agent fake responses that build on each other | Real-target success dropped from 95.8% to 79.2%, and it tricked attackers into revealing their own API keys in 18.8% of runs |
Sources: Gingerbread House, AgentSnare, AgentTrap. Notice the spread: some techniques stop nearly everything, others only slow the attacker down. Deception is a strong early-warning layer, not a replacement for patching and access control.
The catch: AI attackers can learn to spot honeypots
Here is the honest limit. A November 2026 paper, Say Shibboleth, summarizes research on the Rouxii AI pentester: without help, it identified open-source honeypots correctly in 19% of cases. When its prompt included specific honeypot-detection clues, that jumped to 97%. Other researchers have built attack agents that simply skip hosts they suspect are decoys.
So the gullibility is real, but it is not permanent. It lasts until someone tells the AI what a trap looks like. The practical lesson is not “honeypots do not work.” It is do not use bait that looks like bait. Default installs of well-known open-source honeypots have recognizable fingerprints. A file literally called passwords.txt containing admin:admin will be skipped by any agent that has been warned. Bait that blends in with your real environment will not.
| Looks like a trap | Looks real |
|---|---|
| passwords.txt in the root folder | A .env file inside an old project folder, next to real config files |
| admin / Password123 | An AWS key in the exact format AWS uses, named after a real-sounding service |
| A brand-new server with nothing else on it | A decoy that matches your naming scheme, like backup-db-02 |
| A default honeypot install, banners unchanged | Bait tokens placed inside systems you already run |
Set up your first trap for free in 15 minutes
You do not need a big budget to start. Canarytokens, a free service from Thinkst, lets you create bait items that alert you by email when they are used. Its documentation states the tokens are completely free.
Option A: a fake AWS key (best all-round bait)
- Go to canarytokens.org and choose the AWS API Keys token.
- Enter the email address that should receive alerts, plus a note to remind yourself where you will put it (for example “web-01, /var/www/old-app/.env”). The note is what tells you which machine was breached.
- Copy the generated key pair into a file where an attacker would look, using the normal AWS format.
- If anyone uses the key, you get an alert. Per the Canarytokens AWS docs, the alert can take between 2 and 30 minutes because it passes through Amazon’s logging first.
# /var/www/old-app/.env (decoy: every value below is a Canarytoken)
APP_ENV=production
DB_HOST=10.0.4.12
AWS_ACCESS_KEY_ID=<access key from your Canarytoken>
AWS_SECRET_ACCESS_KEY=<secret key from your Canarytoken>
AWS_DEFAULT_REGION=us-east-1Option B: other bait that fits AI attackers
The Canarytokens guide lists many token types. A few that suit AI agents especially well:
- MS Word or Excel document: alerts when opened. Name it something an attacker wants, like “VPN access - IT only.docx”.
- Azure Entra ID login and Azure Login Certificate: fake cloud credentials for Microsoft environments.
- Kubeconfig: a fake Kubernetes access file for container environments.
- MCP Configuration: a fake configuration file for AI tool connections, which is precisely the kind of thing a modern AI agent goes looking for.
- HTTP / web link: a URL that alerts when visited. Use it in decoy text written for AI readers.
Option C: a honeytoken account in Microsoft Defender for Identity
If your company runs Active Directory with Microsoft Defender for Identity, you already own a deception feature. You can tag a dormant account as a honeytoken; per Microsoft’s documentation, any sign-in from a honeytoken account triggers an alert.
- Create a believable account that no one uses, for example
svc-backup-legacy, with a strong password nobody knows. - In the Microsoft Defender portal, go to Settings > Identities > Entity tags, choose Honeytoken, then Tag users and select the account.
- Optionally leave a “forgotten” reference to it somewhere an attacker would find it, such as an old script.
Not sure where Defender for Identity fits next to your other Microsoft tools? Our comparison of Defender for Identity vs Defender for Endpoint explains it.
Where to put bait
| Place | Bait | What it catches |
|---|---|---|
| Web servers and old apps | Fake .env file with a canary AWS key | An attacker who got in through the website and is now hunting for cloud access |
| File shares | Word or Excel document with a tempting name | Someone browsing internal files, including ransomware crews before they encrypt |
| Active Directory | Honeytoken account | Attackers trying stolen or guessed credentials inside the network |
| Code repositories and CI | Fake cloud key or kubeconfig in an old branch | Leaked repos and compromised build pipelines |
| Developer laptops | Fake MCP configuration or SSH config entry | Malware or an AI agent rummaging through a developer machine |
One warning before you scatter bait around. Your own AI tools read files too. A coding assistant indexing a repository, or an internal AI agent with file access, may open a decoy or even follow decoy text meant for attackers. Keep bait out of the folders your own assistants work in, and write down where every token lives so a real alert is never mistaken for a false one. Our guide to non-human identities and AI agent security covers how to keep track of what your own agents can reach.
When a paid deception product makes sense
Free tokens are enough to start and are a great fit for small teams. Paid products add managed decoy servers, central dashboards, and alerting that plugs into your SOC. Two well-known options:
- Thinkst Canary: from the makers of Canarytokens. Decoy devices that imitate real servers and network gear, managed from one console.
- Tracebit: decoy resources (“canaries”) for cloud accounts, CI/CD pipelines and workstations.
Larger security platforms also include deception features, so check what you already pay for before buying something new. Whatever you choose, the rule from earlier still applies: the decoy is only as good as how real it looks.
Want to know what an attacker, human or AI, would find on your website before they get the chance? Run our free website vulnerability scanner.
Frequently asked questions
What is a honeypot in cybersecurity?
A honeypot is a decoy built to be attacked: a fake server, file, account or credential that no legitimate user has any reason to touch. When something interacts with it, you know you have an intruder. Small bait items like fake keys or documents are often called honeytokens or canary tokens.
Do honeypots really work against AI attackers?
Yes, and in some ways better than against humans. AI agents check everything, trust what they read and act fast, so they find and use bait quickly. Lab research in 2026 showed decoys cutting successful AI attacks sharply, in one study from 20 of 40 runs to zero. The limit is that AI agents told to look for honeypots get much better at spotting them, so the bait must look realistic.
Can an AI agent tell that a honeypot is fake?
Sometimes. One AI pentester spotted open-source honeypots 19% of the time on its own, and 97% of the time when given detection hints. Default installs of known honeypot software are the easiest to spot. Bait placed inside your real systems, matching your real naming, is much harder.
Is it legal to put honeypots on my own network?
Placing decoys on systems you own or manage, to detect intruders, is a standard and widely used defensive practice. If you plan to record attacker activity in detail or run decoys that interact heavily with outside parties, check with your legal team about logging and privacy rules where you operate.
How much does it cost to start?
Nothing. Canarytokens are free, and if you already have Microsoft Defender for Identity, honeytoken accounts are included. Paid platforms like Thinkst Canary and Tracebit make sense once you want managed decoy servers and central alerting.
Will a honeypot stop an attack?
Usually not on its own. A honeypot is an alarm, not a lock. Its job is to tell you early that someone is inside, so you can respond before real data is stolen or encrypted. Keep patching, use strong access control, and treat decoys as the tripwire that buys you time.
Reader desk
Discuss this guide
Rate the guide or ask a practical follow-up. Clear questions publish immediately; uncertain submissions wait for review.
Microsoft Cloud Solution Architect
Cloud Solution Architect with deep expertise in Microsoft Azure and a strong background in systems and IT infrastructure. Passionate about cloud technologies, security best practices, and helping organizations modernize their infrastructure.
Share this article
Licensed under CC BY 4.0: you may share and adapt this article with attribution to Protego.

