Cybersecurity · advanced

Detection Engineering for AI Agent Telemetry

A security analytics pipeline for AI systems — synthetic logs of prompts, model responses, retrievals, and tool calls streamed into OpenSearch, with Sigma-style detection rules that fire on jailbreak loops, credentials pasted into prompts, and abnormal retrieval volume. Blue-team skills applied to a brand-new attack surface.

Project overview

Companies are deploying agents faster than their SOCs can monitor them, and almost no detection content exists for AI-specific abuse. Building it puts you in a category of one for AI security roles and shows you can apply classic blue-team method to a new domain.

Real-world use case

A company runs an internal agent that can query customer records. The SOC needs to know within minutes if someone is trying to jailbreak it, exfiltrate data through repeated retrievals, or leak an API key in a prompt.

Tools and technologies

• OpenSearch Security Analytics
• Sigma Rules
• Python (log generator)
• Docker Compose
• MITRE ATLAS

Skills you can demonstrate

Detection Engineering, Log Pipeline Design, AI Threat Modeling, SIEM Operations, Rule Tuning

Build roadmap

1. Stand up OpenSearch
   Docker Compose with OpenSearch and Dashboards. Enable Security Analytics and confirm you can create a detector on a sample index.
2. Design the event schema
   Define fields for each agent event type. Use the AI prompt to pressure-test it — what would you need to investigate an incident after the fact?
3. Build the log generator
   A Python script that emits realistic benign traffic plus scripted attack sequences, with a flag to replay a specific scenario on demand.
4. Write the detection rules
   Implement five rules as custom Security Analytics rules. Start with the credential-in-prompt rule — it is the easiest to get right and the most convincing.
5. Tune against noise
   Run 24 hours of synthetic benign traffic and count false positives per rule. Adjust thresholds and document the trade-off for each.
6. Write the runbooks
   For each detection: what it means, how to confirm it, what to do next. Map each to MITRE ATLAS. This is what a SOC lead will read first.

Ways to extend the project

Add a correlation rule that links a jailbreak attempt to a subsequent spike in retrieval volume from the same session, and map every rule to MITRE ATLAS techniques.

How to showcase your work

• GitHub: The log generator, the OpenSearch setup, the rules, and a runbook for each detection explaining what to do when it fires.
• LinkedIn: Show the dashboard lighting up as the generator replays an attack sequence, with the alert list on the side.
• Resume: Position it as blue-team innovation: "Designed telemetry and detection rules for LLM agent abuse, mapped to MITRE ATLAS, deployed in OpenSearch."