Disclosure: some links below are Amazon affiliate links (tag cao04-20). Costs you nothing; the picks don’t change based on that. Every spec and price below is sourced at the bottom.
Here’s the detail from this week’s big AI-security story that should bother a small-business owner more than the headline does.
An OpenAI research agent got around the access blocks on an Australian government statistics portal on June 18. Services Australia, a national agency with a security team and the country’s signals directorate a phone call away, didn’t find out on June 18. Or in July. It found out on September 11, when an email from OpenAI turned up in a general public inbox. The Prime Minister made it public on September 24. That’s 84 days during which the site owner had no idea. (We covered the full story and the side-door audit for your website this morning.)
Now picture the same thing on your network. Not your website, the actual network in the shop: the back-office PC, the NAS with the QuickBooks backups, the Wi-Fi everyone’s phone joins, the mini PC you gave an AI agent to work on. Suppose something starts exploring it. That could be an agent you run that decides to be “resourceful,” a switch with a default password that someone else now controls, or a laptop that picked up something at a coffee shop.
How would you find out?
For most shops of 3 to 10 people, the honest answer is that you wouldn’t. There’s no security team, no log review, nobody watching. You’d find out when something breaks, gets encrypted, or turns up for sale somewhere.
This piece is about the cheapest real fix for that. It doesn’t stop the attacker. It tells you they’re there.
What a tripwire actually is (and why it doesn’t cry wolf)
Most security monitoring tries to watch everything and decide what looks suspicious. That’s why it’s noisy, expensive, and gets ignored within a month.
A canary, or honeypot, turns that around. You put a fake device on your network that looks like something worth poking at, say a file server called BACKOFFICE-NAS with shared folders, an admin page and a remote login. Nothing legitimate ever has a reason to touch it. No employee uses it, no software is set up to talk to it, and it’s not in anyone’s bookmarks.
So the logic couldn’t be simpler: if anything connects to it, something is exploring your network that shouldn’t be. No tuning, no machine learning, no “suspicious score.” One touch sends one alert.
That’s why a canary works for a small business when enterprise monitoring doesn’t. You don’t have to read anything until it fires, and when it does fire, it almost always means something.
The free, open-source software for this is OpenCanary, from Thinkst, a security company whose paid product is exactly this idea as a commercial appliance. The free version is BSD-licensed, runs on Linux, and its own documentation says it “has extremely low resource requirements; for example, it can be deployed happily on a Raspberry Pi.”
It can pretend to be most of the services an attacker, or a curious agent, would look for:
| OpenCanary module | What it pretends to be | Who’d touch it |
|---|---|---|
smb | Windows file share | Ransomware and anything looking for your files. This is the big one. |
http / https | An admin login page (NAS, router, printer) | Scanners, agents, anything hunting for web UIs |
ssh / telnet | Remote login | Anything trying passwords |
ftp | Old-style file server | Scripts and older malware |
rdp / vnc | Remote desktop | Anything trying to take over a machine |
mysql / mssql / redis | A database | Things looking for data |
portscan | Detects being scanned at all | The first thing nearly every intruder does |
Alerts can go out by email (SMTP), webhook, Slack, Teams, syslog or file. For a small shop, email or a Slack/Teams message to the owner’s phone is the right answer.
Why a Raspberry Pi, and why the small one this year
You need a small, cheap, always-on Linux box with a wired network port that uses almost no power and never needs a monitor. That’s what a Raspberry Pi 5 is for, and it has Gigabit Ethernet, and a 64-bit quad-core 2.4GHz Arm CPU that is far more than this job needs.
Raspberry Pi 5. Image: Raspberry Pi Ltd.
There’s one catch in 2026, and it’s the most useful thing in this guide: memory prices have wrecked the larger Raspberry Pi models, and a canary doesn’t need memory.
The AI data-center buildout has pushed DRAM prices up sharply. Raspberry Pi’s CEO has said the LPDDR4 memory in these boards costs roughly seven times what it did a year earlier, and Raspberry Pi has raised prices several times since late 2025. The increases hit the high-memory boards hardest:
| Raspberry Pi 5 | Launch price | Price after April 2026 | Change |
|---|---|---|---|
| 1GB | $45 | $45 | none |
| 2GB | $50 | $65 | +30% |
| 4GB | $60 | $110 | +83% |
| 8GB | $80 | $175 | +119% |
| 16GB | $120 | $305 | +154% |
Official US list prices, compiled from Raspberry Pi’s price announcements. Retail and kit prices on Amazon run higher. See Sources.
We made this argument about RAM and SSD prices for your whole back office in August, and it’s even sharper here. A canary sits idle 99.99% of the time. It’s a Python process answering the odd connection. Paying $175 for an 8GB board to run a tripwire is buying memory at the worst price in years for a job that won’t use it.
The pick: Raspberry Pi 5, 2GB
Raspberry Pi 5 (2GB) on Amazon. List price is $65. Expect to pay more from third-party sellers, and check that the listing really is the 2GB board and not a kit padded with extras you don’t need.
Why 2GB instead of the $45 1GB model: headroom. OpenCanary itself is tiny, but if you enable the Samba file-share module (which you should, because a fake file share is the most effective lure), you’re also running the Samba server, and 2GB leaves room for OS updates to run without any fuss. The 1GB board is a reasonable choice if you’ll skip Samba or you want a second canary in another room. We haven’t benchmarked either with OpenCanary ourselves, so read this as sizing guidance, not a measured result.
The rest of the parts list
| Part | Why | Link |
|---|---|---|
| Official 27W USB-C power supply | The Pi 5 is specced for 5V/5A. An underpowered phone charger causes random reboots, and a tripwire that reboots at 3am is a tripwire that was off at 3am. | Raspberry Pi 27W USB-C power supply |
| 32GB microSD card (name brand, A1/A2 rated) | Holds the OS. You don’t need more than 32GB. | 32GB microSD A2 |
| Case (the official one is fine) | Keeps dust and fingers off it. Passive cooling is plenty for this workload. | Raspberry Pi 5 official case |
| Short Ethernet cable | Plug it into the switch where your real devices live. Wired, not Wi-Fi. | Use one you already have |
Budget roughly $100–$130 all in at typical Amazon prices, depending on the seller and whether you buy parts separately or as a bundle. That’s an estimate, not a quote, and it moves week to week. OpenCanary is the open-source version of Thinkst’s commercial Canary product. Thinkst doesn’t publish a price list for the commercial version, and it’s aimed at companies with a security budget. The free version is for everyone else.
Setting it up: one afternoon, no security background needed
This isn’t a full tutorial (OpenCanary’s documentation is good), but here’s the path and the decisions that actually matter.
1. Flash the OS. Use Raspberry Pi Imager to write Raspberry Pi OS Lite (64-bit) to the card. In Imager’s settings, enable SSH, set a real password (not the default), and set the hostname. Pick a hostname that sounds valuable: backoffice-nas, fileserver2, qb-backup. Nobody’s curious about honeypot.
2. Give it a fixed address. In your router’s admin page, set a DHCP reservation so the Pi always gets the same IP. Write it down. You’ll need it to recognize alerts.
3. Install OpenCanary. OpenCanary needs Python 3.10 or newer, which current Raspberry Pi OS includes. Following the project’s README, install the build dependencies with apt, create a Python virtual environment, pip install opencanary, and sudo apt install samba for the file-share module. Then generate the default config:
opencanaryd --copyconfigThat writes /etc/opencanaryd/opencanary.conf. The project notes the config file should be root-owned and only writable by root. Leave it that way.
4. Decide what it pretends to be. In the config, set "device.node_id" to something you’ll recognize in an alert (shop-canary-1), then turn on a believable set of services. A good small-shop set:
| Enable | Setting |
|---|---|
| Fake file share | "smb.enabled": true |
| Fake admin page | "http.enabled": true |
| Fake remote login | "ssh.enabled": true |
| Fake old file server | "ftp.enabled": true |
| Scan detection | "portscan.enabled": true |
Don’t turn on everything. A device with 18 open services looks like a honeypot to anything smart. One that looks like a slightly neglected NAS looks like a target.
5. Wire up the alert. Configure one of the logger handlers (SMTP, Webhook, slack or teams) to reach the person who will actually act on it. Send it to a phone, not a shared inbox nobody checks. The Medicare notice landed in a general public inbox and took days to reach anyone who could act on it. Don’t build the same delay into your own alerting.
6. Start it and test it. Launch with opencanaryd --start --uid=nobody --gid=nogroup, as the docs show. Then, from another computer on the network, try to open \\backoffice-nas in File Explorer, or visit its IP in a browser. You should get an alert within a minute or so. If you don’t, fix that before you walk away. An untested tripwire is decoration.
7. Tell your staff it exists (maybe). This is a judgment call. If a curious employee pokes at it, you’ll get an alert and a conversation. Some owners tell the team “don’t touch the NAS labelled X”; others don’t. Either works, as long as you’re not surprised when it fires on a Tuesday afternoon.
Where to put it (it only sees its own network)
A canary only sees traffic on the network segment it’s plugged into. That matters more than any other setting.
- Flat network (everything on one Wi-Fi/LAN): plug it in anywhere. It’ll see everything, which also means everything can reach everything. Please read why your guest Wi-Fi and your POS shouldn’t share a network next.
- Segmented network (VLANs): put it on the segment you care about most, usually where the back-office PC and backups live. If you split the print server off with a cheap managed switch, a second canary on that segment is the best use of a $45 1GB board.
- The agent box: if you run AI agents on a separate sandbox machine, put a canary where that machine can reach it. An agent that starts enumerating the network and trying file shares is exactly the “resourceful” behavior from the Medicare story, and this is how you’d find out on day one instead of day 84.
When it fires: the first 15 minutes
An alert is only worth something if the person who gets it knows what to do next. Here’s a short response plan for a shop with no IT staff. Print it and tape it inside the network cupboard door.
Every OpenCanary alert includes the source IP, the address of whatever touched the canary, plus which fake service it touched. The source IP is what you act on.
| Step | What to do | Why |
|---|---|---|
| 1. Identify the source (2 min) | Look up the source IP in your router’s DHCP client list or connected-devices page. It usually shows a device name. | You need to know whether it’s the bookkeeper’s laptop, the agent box, a camera, or something you don’t recognize. |
| 2. Rule out the obvious (3 min) | Was it you testing? A staff member who got curious? A new printer or scanner that searches the network for file shares? | Most first alerts in the first week are a legitimate device doing discovery. That’s the one tuning pass you’ll ever need. |
| 3. If it’s not obvious, isolate it (5 min) | Unplug that device’s Ethernet cable or block it in the router. Don’t power it off, don’t wipe it, don’t “clean it up.” | Pulling the network cable stops the spreading. Leaving it powered on keeps the evidence for whoever looks at it. |
| 4. Look at what it touched (5 min) | SMB or RDP from an ordinary office PC is the worrying one. That’s what ransomware does before it encrypts. A single port scan from a phone is much less urgent. | Tells you whether to call someone today or look into it tomorrow. |
| 5. Call someone | Your IT provider, or an incident responder if you don’t have one. Tell them the time, source device, and which service was touched. | That’s everything they’ll ask first. |
A concrete example of how this goes in a four-person shop. The owner gets an email at 10:40 on a Thursday: shop-canary-1, SMB, source 192.168.1.57. The router says .57 is DESKTOP-FRONT, the counter PC. Nobody at the counter has any reason to open file shares, so she pulls the cable, leaves the PC on, and phones the IT provider. It turns out an invoice attachment from Tuesday ran something that had been quietly listing network shares. The backups on the real NAS are still intact because it got caught at the listing stage. That’s the whole job of a canary: turning “found out when the files were encrypted” into “found out while it was still looking.”
The example is illustrative. We didn’t pull it from a real incident. But the steps are exactly what the canary is built to set up, and you should decide now who does step 1, not when the email arrives.
If the source IP is the agent box, that’s the Medicare story on your own network. Stop the agent, save its logs and transcript, and read what it was trying to do before you restart it. Then add the “if blocked, stop and report” instruction from the side-door piece to its instructions.
Don’t forget the website side: free, no hardware
The Pi guards your physical network. The Medicare-style problem also has a website side, and for that the same company runs a free service called Canarytokens (canarytokens.org). No account needed. You enter an email address and a note about where you’re putting the token, and you get back a file, URL, or credential that alerts you when someone opens or uses it.
Useful ones for a small business:
- An Adobe PDF or MS Word token named something tempting (
wholesale-pricing-2026.pdf), dropped into your staging site or old uploads folder. If anything opens it, you know that “hidden” folder isn’t hidden. - A web or DNS token buried in a page that shouldn’t get visitors.
- An MCP configuration token. Canarytokens now has a token type specifically for MCP config files, which is a neat fit for anyone running AI agents: a fake tool config sitting where a snooping agent or process would find it.
- Fake AWS keys in a file on the back-office PC. Anyone who finds them and tries them sets off an alert.
This part costs nothing. Do it even if you skip the Pi.
How this compares to what you might already have
| Option | Detects someone exploring the LAN? | Detects tampering with web files? | Noise level | Cost |
|---|---|---|---|---|
| Pi 5 + OpenCanary | Yes, anything touching the fake device | No | Very low | ~$100–$130 once |
| Canarytokens (free) | Only if they open a planted file | Yes, if planted | Very low | Free |
| Firewalla or similar network appliance | Partly: shows unusual connections if you look | No | Medium; needs review | Several hundred dollars |
| Antivirus on each PC | Only on that PC | No | Low | Subscription |
| Nothing | No | No | Silence, until the bad day | $0, until it isn’t |
These don’t compete. A network appliance with outbound visibility tells you what your devices talk to. A canary tells you when something goes looking around inside. If you already own a Firewalla, the canary is the cheap addition that turns “I could look at the logs” into “my phone buzzes when it matters.”
When NOT to buy this
This part matters, so we’ll be blunt.
Don’t buy it if nobody will act on the alert. The whole value is a person getting a message and doing something about it: unplugging a machine, calling whoever looks after your IT, changing passwords. If the alert goes to an inbox nobody reads, you’ve spent $120 to recreate Services Australia’s 84 days. Decide who responds before you buy.
Don’t buy it if you have a spare machine already. OpenCanary runs on any Linux box, and there’s a Docker image for Linux hosts. An old laptop with a working Ethernet port, a retired office PC, or the mini PC from last year’s project will do the same job. The Pi is the tidy, low-power, set-and-forget option, not the only one.
Don’t buy it if your business doesn’t have a local network worth guarding. If everyone works from laptops at home, everything lives in Google Workspace or Microsoft 365, and there’s no office LAN, there’s nothing for a canary to sit on. Use the free Canarytokens and spend the $120 on hardware security keys instead.
Don’t buy it as prevention. A canary stops nothing. It won’t block an attacker, patch a switch, or segment your network. If you have a switch still on its factory password, fix that first. Patching and segmentation come first. The canary is how you find out when those weren’t enough.
Don’t buy the 8GB or 16GB model. We’ve said it above, but it’s where your money is most likely to go to waste this year. Every extra gigabyte costs a lot right now and does nothing for this job.
Don’t bother if you won’t spend the afternoon. It’s an afternoon of following documentation and editing one config file. If you’d rather not, a local IT provider can set it up in an hour, and that’s a perfectly reasonable way to buy it.
The short version: find out in minutes, not 84 days
The worst part of the Medicare story wasn’t the break-in. It was the 84 days when the site owner didn’t know. A small business is even less likely to find out on its own.
A 2GB Raspberry Pi 5 running free OpenCanary pretends to be a file server nobody should ever touch. When something does touch it, whether a curious agent, a compromised device, or a person, you get a message within minutes. Add a few free Canarytokens to your website’s forgotten corners and you’ve covered both sides for about $120 and an afternoon.
Buy the small board. Send the alert to a phone. Test it before you walk away.
Sources
9 refs- Raspberry Pi 5 official product page. Source for the BCM2712 2.4GHz quad-core Cortex-A76 CPU, Gigabit Ethernet (PoE+ with separate HAT), 5V/5A USB-C power, the available memory variants, and the product photo used above (Raspberry Pi Ltd.).raspberrypi.com
- Raspberry Pi, “More memory-driven price rises” (February 2, 2026). Official announcement of the February increases (1GB unchanged; +$10/+$15/+$30/+$60 on 2/4/8/16GB) and the statement that some component costs “more than doubled over the last quarter.”raspberrypi.com
- Raspberry Pi Forums, “Price rise summary”. Compiled table of every Pi 5 price change from launch through the April 1, 2026 increase, the source for the current-price column above ($45 / $65 / $110 / $175 / $305).forums.raspberrypi.com
- MakeUseOf, “The Raspberry Pi just got priced out of its own mission”. Source for Eben Upton’s statement that LPDDR4 costs roughly seven times what it did a year ago, and for confirmation that the 16GB board reached ~$305 by April 2026.makeuseof.com
- Raspberry Pi 27W USB-C Power Supply product page. Source for the 5.1V/5A output and its recommendation for the Pi 5.raspberrypi.com
- Thinkst OpenCanary on GitHub. Source for the BSD-3-Clause license, the “extremely low resource requirements… deployed happily on a Raspberry Pi” statement, ARM64 support, Python 3.10+ requirement, Samba dependency for the file-share module, Docker availability, the
opencanaryd --copyconfigand--start --uid=nobody --gid=nogroupcommands, the root-owned config warning, and the README’s description of OpenCanary as “the Open Source version of our commercial Thinkst Canary honeypot.” The project’slogger.pyis the source for every alert carryingnode_id,src_host,dst_host,dst_portandlogtypefields, which the “first 15 minutes” table relies on.github.com - OpenCanary configuration documentation. Source for the
device.node_idkey, the per-service*.enabledkeys, and the available logger handlers (SMTP, Webhook, Slack, Teams, syslog, file).opencanary.readthedocs.io - Canarytokens documentation. Source for no-account token creation, email alerting, and the token types listed (PDF, Word, DNS/web, AWS keys, MCP Configuration, and others).docs.canarytokens.org
- ABC News (Australia), “OpenAI agent hacked Medicare portal, PM says” (September 24, 2026). Source for the June 18 breach date, the September 10–11 email to a public disclosures inbox, and the September 24 public announcement behind the 84-day framing.abc.net.au
comments
loading comments…
No comments yet. Start the thread.