What the crews did with my decoy: seven days on a honeypot

November 14, 2026. This is the follow-up to How my booking page got popped, and what I changed. That post ends with a question I could not answer: once the attackers had a password that could read my mail, did they read it? I built an instrument to find out. This is what it saw in seven days.

What I set up

Short version: a decoy. A second copy of the same booking app, with the same known bug deliberately left in, on a throwaway cloud server that had nothing of mine on it and no way to reach anything of mine. Then I waited for crews using the same bug to find it, and watched what they did with what they found.

The parts, in plain terms:

  • The lure. The vulnerable image, byte for byte the one that had been exploited on my real booking page, running as a fictional “Appointments Desk” on its own hostname. Sign-ups off, telemetry off, one fictional admin user.
  • The bait. The app’s configuration held twenty settings shaped exactly like the real ones, with none of the real values: a decoy mail login for a decoy mailbox on a decoy mail server I ran for the purpose, random database and signing keys, a fake Google OAuth client that matches nothing, and a canary: an Amazon cloud access key that belongs to a tripwire service (Canarytokens, free from Thinkst), not to me. Use that key anywhere in the world and I get told where from.
  • The cage. The server could not talk to the internet at all, except to the two addresses the original malware had used to fetch its payload and send its loot, and even those were watched. No mail could leave it. Nothing on it pointed at anything I own.
  • The instruments. Every command the app ran, every connection in or out, every DNS lookup, every login attempt on the decoy mail server, and the web proxy’s log, pulled off the box every five minutes to a machine at home. A watcher compared the box against a baseline on each pull and had a kill switch that could cut the server off at the network edge in seconds if anything looked like it was escaping the cage.
  • The clock. Seven days from the moment the server was built on November 7, then an automatic teardown that revoked the decoy credentials and destroyed the servers. That happened this afternoon. The lure itself, after reviews and two aborted publishes, was live for the last four of those days.

Everything above was reviewed by a second, independent model before each step, and the first two attempts to publish the lure were stopped by my own kill switch because a check failed. I would rather be killed by my own safety than leak something real. The lure finally went live on the evening of November 10 and, after one false alarm from the watcher took it down for an hour, stayed up until the teardown.

What arrived

The internet found it without any help from me. In the first hour the bare address, before it even had a name, took 974 connection attempts from scanners. Once the name resolved, the traffic sorted itself into four kinds.

Scanners. The background noise of the public internet: probes for .env files, WordPress APIs, old PHP bugs, Exchange, Jira, git directories. By Friday morning the lure had seen about 2,650 outside requests from more than 316 addresses, and 87 of them were just for .env. One security-research scanner found the bug within six hours of go-live, confirmed it with a harmless check, and presumably listed the host on its public index, which likely brought the next visitors.

The harvester. Nine hours after go-live, a single carefully built request from an address on Scaleway’s network exploited the bug and ran a reconnaissance script inside the app. It returned a compact report: what the machine was, and every environment variable the app process could see, 41 of them, including the decoy mail password. It did not drop malware, did not open a shell, did not come back. It was a harvester: take the credentials and move on. Over the week the same kit, recognisable by its request shape and its Python HTTP library, ran from five more addresses on hosting networks in the United States, Germany, Indonesia and Singapore, 141 times in all. By the last day it was running every few minutes.

The credential hunter. The most interesting visitor. Sixteen hours after go-live, a different crew used the same bug to run real shell commands on the server, 36 of them in five seconds. The commands tell you what they wanted: dump the environment, read every file in the app directory, and then, specifically, go looking for the credential files of AI coding agents: the OpenAI Codex CLI’s token file and Claude Code’s settings file. Those files did not exist on the lure. The crew came back on a schedule, every two and a half hours, seventeen times over three days, running the identical kit each time. Then it stopped. Each run read the decoy mail password again. By the end it had read it 272 times without once using it.

A fourth kit, from a different European hosting range, arrived on the Friday with a broader shopping list: shell histories, Kubernetes service tokens, cloud credentials, the environment. It read the decoy environment eight times. A fifth, on the final day, read it four more times.

The mail server. The decoy mailbox sat on its own server, reachable from anywhere, logging every connection. It got what every mail server gets: TLS cipher scans running through every protocol version from SSL 3 up, a relay probe that tried to send mail through it without logging in, three attempts to log in with a plaintext password that the server refused, and a handful of scanners that announce themselves by name before they even say hello. None of that touched the decoy account.

Where it came from

Every address that exploited the bug or tried the credentials was registered to a hosting or cloud network rather than a residential provider: Scaleway in France, DigitalOcean in the United States, Contabo in Germany, a Utah-based VPS reseller, Tencent Cloud in Singapore, a fibre provider in Indonesia, and smaller ranges in Romania, Bulgaria, Denmark and the Caribbean. Network ownership says where a machine sits, not who rents it or where the operator is. I am not naming individual addresses: they are rented, they rotate, and the next tenant is not the attacker.

Did they use the password?

This is the question the whole instrument was built to answer, so here is the careful version.

For 87 hours, no. Five different kits stole the decoy mail password a combined 425 times, and in that time the instruments recorded no login to the mailbox and no attempt to send mail with it. I had a positive control for the login path: in the middle of the window I logged in to the decoy mailbox myself from home, and the instruments recorded that login correctly. So the zero is a zero on a path that was proven to record logins, with the usual caveat that a control proves the path it exercised, not every path all week.

Then, twenty minutes before the teardown, yes. At 3:10 pm Pacific on November 14, something logged in to the decoy mail server over SMTP with the stolen password, from an address on a Danish hosting network that none of the earlier visitors had used. It authenticated successfully and sent nothing. That pattern is consistent with a credential checker, a script that confirms which stolen logins still work; whether the list was sold, shared or simply re-run by the same operator from a new machine, I cannot tell from one login. What I can say is that 87 hours after the first theft the password was tried from somewhere new, and that nothing ever tried to read the mailbox over IMAP.

And the canary sang. In the same half hour, the tripwire cloud key that had been sitting in the stolen environment all week was reported used twelve times by the tripwire service, from two hosting ranges in Utah and Missouri. At the same moment the lure itself resolved the names of Amazon’s token service, its secrets manager, its storage service, and Bedrock, Amazon’s hosted AI-model API. A name lookup is not an API call, and the tripwire tells me the key was presented, not what was asked; but together they say someone took the key from the stolen environment and tried it, from the lure and from their own machines, with AI-model access among the first things on the list. The key can do nothing; it exists to report that it was tried.

Put together: the crews working this bug harvest credentials wholesale, and the credentials move. A mail password sat unused for three and a half days and then was tried from a network none of the harvesters had used. A cloud key was tried the same afternoon, with AI services among the names looked up. Nothing read the decoy mailbox.

What it cannot tell me

The honest caveat. My real credential was a Gmail app password; the decoy was a login on a small self-hosted mail server. An SMTP check is comparable between the two, and that is the part the decoy caught. But a Gmail credential is a far more valuable prize, the kind that invites someone to download the whole mailbox over IMAP, and a self-hosted decoy is a weaker lure for that specific behaviour. So “nobody read the decoy mailbox” is evidence, not proof, about my own. The way to answer the IMAP question properly would be a decoy Gmail account, which needs a phone number to create and is a decision for another day.

I also only saw what came through the front door. The crews could have shared the harvested environment with each other out of my sight; the Danish login and the canary hits suggest exactly that. And seven days is seven days. The crew that returned seventeen times on a schedule stopped on day four; I do not know why.

What I learned

  • Exploitation is industrial, and the product is credentials. Nobody defaced anything, nobody mined, nobody persisted. They read the environment and left. The harvesting kits ran unattended, on schedules, from rented machines, and the same bug was hit by at least six distinct toolkits in a week.
  • AI-agent credentials are now on the shopping list. One kit’s first move after the environment dump was to look for Codex and Claude Code token files. If you run coding agents on a server, their credential files are a target with a name.
  • Stolen credentials are verified later, elsewhere. Three and a half days of silence meant nothing. The check came from a network that had never touched the lure.
  • Canary keys work. A fake cloud key cost nothing, sat quietly for a week, and then told me the minute and the networks from which someone tried it, with AI services among the names the lure looked up at the same moment.
  • Build tooling can be a defence by accident. Six of the twenty settings, including the cloud key and the database password, were never captured by the in-process harvester, because the app’s build system strips undeclared environment variables from the server process. The shell-command kit got them anyway, by reading the process that launches the app. Defence in depth is real, and so are its gaps.
  • Fail-closed instruments have a cost. My watcher killed the lure once on a transient network hiccup. I took the false positive, fixed the retry, and put it back. I would make the same trade again.

What I changed because of it

The booking-page post lists the changes the breach itself forced: least privilege for mail, a build gate that refuses images with known-exploited bugs, a deploy gate, per-finding triage, and a watch on my own Sent folder. The decoy added these:

  • No agent credential files on any server. Coding-agent tokens live on my workstation and in the password manager, never in a home directory on a host that serves the public.
  • Canary keys wherever they are cheap. Tripwire credentials now sit in places a harvester would read. A quiet canary costs nothing; a loud one is the earliest warning I have.
  • Treat harvested-but-unused as used. If an environment is ever dumped, every credential in it is rotated the same day, whether or not anything has been seen using it. The 87-hour delay is exactly the window in which “nothing happened yet” is the wrong conclusion.
  • Checks that print one line, on a timer. The week taught me how much I was reading by hand. The nightly scans, the certificate expiries, the edge load, the password-manager state: each now has a scheduled check that prints one line with a number, a time and a verdict, and a page that goes amber when a line is stale. The instrument is only useful if someone, or something, reads it on time.
  • A documented, rehearsed teardown. The decoy’s credentials were revoked and its servers destroyed by a timer, with proof read back, not by someone remembering to do it. That is now the standard for anything I stand up on purpose to be attacked.

Seven days of cage, four of lure, six toolkits, 425 thefts of a password that opened nothing, one login from Denmark, and a canary that sang. I did not get my answer about my own mailbox, and I probably never will. I did get a very clear picture of who was on the other side of that bug, what they wanted, and how long the quiet lasts before the stolen thing moves. That was worth a week and a few dollars of rented cloud.

Comments are closed.

This site uses Akismet to reduce spam. Learn how your comment data is processed.