AWS Builder Center

Weekend Creative Agent Challenge: OT/ICS Threat Storyteller

This article is my submission into the Weekend Creative Agent Challenge.

Weekend Creative Agent Challenge: OT/ICS Threat Storyteller

Tags: #agents

Vision and what it does

I spend my working life thinking about defensive cyber operations at the strategic level at the Department of Defense, and outside of work I run a small OT/ICS (operational technology / industrial control systems) homelab. It's a three-zone Purdue Model network built to practice the kind of segmentation and detection work that matters in real critical-infrastructure environments (in-debugging and documentation). One thing I've noticed teaching and building awareness material in this space: a lot of ICS security content is dry. Accurate, but dry. People don't often remember a checklist the way they remember a story.
So for this challenge, I built the OT/ICS Threat Storyteller. It's an always-on agent that wakes up once a day, with no human involvement, and invents a short, creative story dramatizing a plausible ICS attack scenario—like a compromised HMI, a tampered safety instrumented system setpoint, or a ransomware pivot through a flat IT/OT network. It then illustrates the scenario with a generated image (not currently active, will troubleshoot further) and publishes everything to a small public gallery site.
Each story ends with a "Defender's Takeaway" — three concrete, practical mitigations tied to that specific scenario. It's a security-awareness tool wearing a creative-writing costume: something someone might actually want to read on purpose, that still teaches the same lessons a formal advisory would.
The agent draws from a rotating bank of sixteen scenario seeds spanning Purdue Model Levels 0 through 5 — field instrumentation, PLCs and RTUs, HMIs, historians, the IT/OT DMZ, safety systems, even building automation as a pivot point — so it works through real variety before anything repeats.

How I built it

I kept the architecture deliberately simple given the weekend timeline: one Lambda function, triggered daily, doing all the work in a single invocation. The interesting design decisions were less about infrastructure and more about getting reliable structured output from a generative model running completely unattended. When there is no human in the loop to notice or fix a malformed response, the agent has to be its own quality gate.
I addressed that by having the model return a single strict JSON object (title, story, defender's-takeaway lessons, and an image prompt derived from its own story) with a defensive parser on the Lambda side that strips stray markdown fencing and falls back to regex extraction if the model doesn't return pure JSON.
The other real decision was scenario rotation: instead of letting the model free-associate a theme every day (which drifts toward repetition fast), I maintain a small shuffled queue in S3 that works through the entire scenario bank before anything repeats. This keeps the archive varied over time instead of converging on the same two or three attack patterns.

AWS services used and architecture

  • Amazon Bedrock (Amazon Nova Lite) — generates the story, title, and defender's-takeaway lessons as structured JSON
  • Amazon Bedrock (Amazon Nova Canvas) — generates a themed illustration from a scene description the story model writes for itself (this is the part that is down - cannot seem to find an active Nova Canvas)
  • AWS Lambda — the whole agent's logic: pick a scenario, call both models, write results to S3, regenerate the gallery page
  • Amazon S3 — stores every day's story/image, plus hosts the public site itself via S3 static website hosting
  • Amazon EventBridge — a daily scheduled rule that invokes the Lambda with zero manual triggering
Flow: EventBridge → Lambda → Bedrock (Nova Lite, then Nova Canvas) → S3 (story JSON + image + regenerated  index.html ). No servers, no containers, nothing running except the ~30 seconds a day the function is actually executing.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article