Talk · DevOpsDays Barcelona 2026 · Ignite
One Fake Password Is Worth More Than 10,000 Alerts
Five minutes, twenty slides, two incidents ten months apart. React2Shell, December 2025: the alerts fire and it takes four days to contain. PHP2Shell, October 2026: a fake AWS key trips, an agent restarts the compromised container, and it's contained in under two minutes with zero harm.
Both incidents anonymised and simplified. Views are my own. Coincidences are accidental.
- When
- Sat 14 Nov 2026, 13:15–13:20
- Where
- BAU, Pujades Track, Barcelona
- Format
- Ignite: 20 slides × 15 s
- Speaker
- Ilia Dubovskii
The replay
Watch a decoy get used, and the machine respond
Two minutes on a real AWS account, filmed on 8 November 2026. An attacker's shell inside the app container finds two decoy credentials and tries them. Slack has the alert in seconds, Compass stops the compromised task and switches the key off, and you can follow all of it in the AWS console. The break-in itself is skipped: the replay starts inside the container.
0:05 · Inside the container
The attacker dumps the environment and uses the backup AWS key and the database password. Both are decoys.
0:32 · The alert
3.5 seconds after AWS logs the key's call, the alert is in Slack and Compass has opened the incident.
0:42 · The task stops
In the ECS console the attacker's task is stopped, with the Compass incident as the reason, and a clean copy starts.
1:20 · Contained
Verified 66 seconds after first use; the stolen key is switched off. Then AWS's own CloudTrail record of the call.
How it was filmed: every screen is a live recording of the same run, with one clock. Waits are sped up four times and marked "×4". Secrets and the decoy key ID are masked. Across nine runs that day, the median was 4.2 s to the alert and 63 s to containment.
The talk in four lines
An alert is a maybe.
CPU, traffic and cost all have innocent explanations. Ten thousand of them a month teach people to wait.
A honeytoken is a certainty.
Nobody legitimate uses a fake key. When it's used, you know it's an attacker, and exactly which task leaked it.
Automate on certainty.
AI acting on maybes is dangerous. AI acting on certainty, with an action you can undo, is good engineering.
Keep the alerts.
Use them as context. The honeytoken decides; the alerts explain.
Slides and transcript
All 20 slides with what is said on each. The full argument is in the paper: Automate on Certainty: Honeytokens as a High-Confidence Trigger for Automated Response in Cloud Environments (preprint, DOI 10.5281/zenodo.23605210).
Try it on Monday
A honeytoken AWS key in plain Terraform: a deny-all IAM user that looks like a real workload, its key, and an EventBridge rule that fires on any use of it, read-only calls included. Plant the key where an attacker would look: container environments, CI variables, a wiki page.
resource "aws_iam_user" "decoy" {
name = "svc-backup-reporting" # looks like a real workload
path = "/service/"
}
resource "aws_iam_user_policy" "deny_all" {
user = aws_iam_user.decoy.name
policy = jsonencode({
Version = "2012-10-17"
Statement = [{ Effect = "Deny", Action = "*", Resource = "*" }]
})
}
resource "aws_iam_access_key" "decoy" {
user = aws_iam_user.decoy.name # plant this key where attackers look
}
resource "aws_cloudwatch_event_rule" "decoy_used" {
name = "backup-reporting-activity"
state = "ENABLED_WITH_ALL_CLOUDTRAIL_MANAGEMENT_EVENTS" # include read-only calls
event_pattern = jsonencode({
"detail-type" = ["AWS API Call via CloudTrail"]
detail = { userIdentity = { userName = [aws_iam_user.decoy.name] } }
})
}
# + an aws_cloudwatch_event_target to your alerting / response (SNS, SQS, Lambda…)
Want this planted, correlated to the exact workload and acted on automatically? That's what Compass honeypots do, and the code is open source (Apache 2.0).
Choosing a honeytoken, or wondering what it adds to GuardDuty? See Thinkst Canary vs Canarytokens.org vs self-hosted and Honeytokens vs GuardDuty.
The talk ends with "bring on Java2Shell". Three days later it had a CVE. Every one so far, with versions, fixes and what to do in the window before the patch: the *2Shell vulnerabilities, Java2Shell, React2Shell, PHP2Shell, Node2Shell and Py2Shell.
Saw the talk?
This is the page the QR code on the last slide points to. Questions are welcome.
Book 30 minutes with Ilia →
Beamreach