DEFCON TORONTO

AI × Cyber Defense

Supercharging Cyberdefense

Kc Udonsi · @glitchnsec

$ whoami

About Me

Security Architect @ Stan
Security Research & Engineering
Building "paved road" security — empowering & protecting innovation
Most dangerous thing I do? ↗

(threat model: myself)

Industry Contributions

Recent Work

FortyOne-OSS — Open source multi-tenant extensible AI OS platform

glitchnsec.github.io/fortyone-oss

DMI — Messaging intelligence service

dmi_web.securecybernaut.com

📚

No Starch Press — Technical Reviewer

If you're writing, hit me up 😉

Blogs & More

glitchnsec.github.io/blogs

GO GRAB A COPY 😉

Black Hat Bash

Dolev Farhi & Nick Aleks

From Day Zero to Zero Day

Eugene Lim

Key Insights

Takeaways

01

AI Agents DIY

Understand and build AI agents from scratch. Understand their strengths and failure modes

02

Rethinking Security Defense for AI

New threat models demand new defensive postures.

03

Design Decisions & the Role of Humans

Avoid cognitive debt & atrophy — keep humans sharp in the loop.

I

Section 01

Part I

AI Agents

What they are, why they matter, and how to build one before lunch.

🙋 Raise your hand if you've used an AI agent this week.

🤔 Now keep it raised if you could explain how it works.

Let's fix that.

You've seen these

Agents in the Wild

🧑‍💻

Claude Code

Terminal-native coding agent. Reads your repo, plans changes, writes & tests code.

📦

Codex

OpenAI's cloud agent. Spins up a sandbox, runs tasks, returns a PR.

🔮

Gemini

Google's multimodal agent. Searches, reasons, calls APIs across the ecosystem.

Different wrappers. Same underlying pattern. →

The Pattern

Agent = ( Model, Memory, Tools, Planning )

🧠

Model

The LLM doing the reasoning. Claude, GPT, Gemini — pick your brain.

💾

Memory

Conversation history passed as messages. Context is everything.

🔧

Tools

Functions the agent can call. Search, code exec, APIs — its hands.

🗺️

Planning

Reasoning config on the model. How hard should it think?

while goal not reached: observe → reason → act → repeat

↑ This is the ReAct loop. That's it. That's the pattern.

DIY

Build One in ~20 Lines

python# pip install -U agno openai openrouter

import os
from agno.agent import Agent
from agno.models.openrouter import OpenRouterResponses
from agno.tools.hackernews import HackerNewsTools

agent = Agent(
    model=OpenRouterResponses(              🧠 Model
        id="anthropic/claude-sonnet-4.6",
        api_key=os.getenv("OPENROUTER_API_KEY"),
        reasoning={"enabled": True}    🗺️ Planning
    ),  # costs more tokens
    tools=[HackerNewsTools()],               🔧 Tools
    markdown=True,
)

# Print the response in the terminal
response = agent.run(                       ⚡ ReAct Loop
    "What's latest on HackerNews?"
    " Related to CVEs"
)
print(response.content)                    💾 Response

Documentation

Agno:
docs.agno.com
OpenRouter
https://openrouter.ai/docs/quickstart

You just built an agent.

seriously, that's it 🎉

Part I — Live

▶

Let's Play

> swapping prompts, tools, memory, reasoning_

I

Still Section 01

Failure modes

Agents are easy to build.
They're also easy to manipulate.

Threat Surface

What Can Go Wrong

🌀

Hallucinations

The model invents facts — or worse, hallucinates tool calls it was never given. Confident nonsense at machine speed.

Build around it

Observability — log & trace every step of the ReAct loop

Proof of work — require evidence before trusting output

🔓

Data Privacy & Residency

Agents ingest everything you feed them. Prompts, context, tool results — all hitting external APIs.

Build around it

PII removal — strip before it leaves your perimeter

Pseudo-anonymization — keep utility, lose identity

💉

Input Sanitization

Prompt injection, malformed tool responses, adversarial user input — classic injection, new surface.

Build around it

Enforce validation — post tool calls and on user input

Schema contracts — never trust unstructured returns

II

Section 02

Part II

Rethinking Defense

Cyber defense has not kept up with offense
when it comes to AI adoption.

The Gap

Offense vs Defense

AI-generated phishing	vs	AI content analysis…?
AI vuln discovery	vs	AI triage & fix…?
AI pentests / red team	vs	AI blue team…?

🤷 Why is defense always a step behind?

Root Cause

Two Reasons

🏛️

Red Tape + Data Privacy

Compliance, legal review, data residency requirements. Attackers don't file change requests.

🧩

Technical Know-How

Skill gap. Security teams know defense, not ML pipelines. AI teams know models, not threat landscapes.

💬 Which one bites your team harder?

Adoption Spectrum

Three Ways to Incorporate AI

Assisted 🧑‍💻 ← 🤖

Human drives, AI helps.

Copilot suggestions, auto-complete, draft generation.

Augmented 🧑‍💻 ⇄ 🤖

Both human and AI own work.

e.g AI triages alerts, drafts rules, human approves.

Native 🤖 → 🧑‍💻?

Rethink from first principles.

Evaluate the entire domain against AI strengths.

First two unlock speed. The third requires rethinking the problem.

Building at the Intersection

I've Been Here a While

DMI WEB

An AI-powered detection & monitoring intelligence platform.

dmi_web.securecybernaut.com

And a few months ago — NebulaFog Hackathon

Where I explored the idea of the autonomous AI SOC.

NebulaFog Hackathon

The Framework

AI Rethink

Concisely describe the problem / success metric

Understand & articulate the workflow of solving it

Understand why each step is relevant ↳ Identify what compensates for or contributes to human fatigue & limits

Measure against agentic strengths / weaknesses

Replace | Insert AI agents — or eliminate the step entirely

Profit 📈

🛡️ Let's build this together — what does a blue team actually do?

The Defense Workflow

Understand the environment

Gather threat intel & research relevant to the environment

Detection engineering write, test, tune, manage… 😮‍💨

Monitor detection events

Threat hunt

(Optionally) Incident response 🤞 hopefully not

Combinatorics

The 6 × 4 Matrix

WORKFLOW STEP	ASSIST	AUGMENT	NATIVE	ELIMINATE
Understand env	—	—	—	—
Threat intel	—	—	—	—
Detection eng.	—	—	—	—
Monitor events	—	—	—	—
Threat hunt	—	—	—	—
Incident resp.	—	—	—	—

4⁶ = 4,096 possible combinations

Too many to evaluate all — but I'll show you one.

One Combination

The 7 × 4 Matrix

WORKFLOW STEP	ASSIST	AUGMENT	NATIVE	ELIMINATE
Understand env	—	🤖 Augment	—	—
Threat intel	—	🤖 Augment	—	—
Detection eng.	—	—	—	✕ Eliminate
Monitor events	—	—	—	✕ Eliminate
Threat hunt	—	—	🤖 Native	—
Incident resp.	—	🤖 Augment	—	—
Compaction NEW	—	—	🤖 Native	—

4⁷ = 16,384 possible combos — this is just one

Back to Part I

Every 🤖 Needs

💾

Context / Memory

Environment data, past findings, conversation state

🔧

Tooling

SIEM queries, enrichment APIs, ticketing, code exec

🏗️

Harness

Agent framework, guardrails, validation, observability

🧠

Model Selection

Right model for the task — cost, speed, reasoning depth

Sound familiar? It's the same tuple from Part I — now applied to defense.

III

Section 03

Part III

Design & Humans

We know where to place agents.
Now: how they work, what humans do, and the code.

Augment

Understand the Environment

🧑 Human

Log curation & ingestion to SIEM

→

🤖 Agent

Extrapolate environment context & nuances

→

🧑 Human

Validate / review / approve environment context

Augment

Cyber Threat Intel & Research

🧑 Human

Curate & validate intelligence feeds

→

🤖 Agent

Extrapolate & research artifacts relevant to env. Theorize potential threats.

→

🧑 Human

Validate / review / approve intelligence artifacts

Eliminate

Detection Engineering

🫣 Don't hate me on this one.

Why eliminate?

DE in its current form is fragile. Over-tuned to reduce FPs. Context & correlation are an active pain point. Massive maintenance overhead. Always a question of coverage in an unbounded domain. Sometimes no room to detect novel attacks.

Substitute with

Security context / summarization / compaction — let agents tell the story instead of writing fragile rules.

Eliminate

Monitoring Events

Why eliminate?

Tedious. Effectiveness varies on context, experience, readiness. DE is optimized for SOC analysts — eliminate DE → eliminate SOC as we know it.

Substitute with

Threat hunting only — skip the alert queue entirely, go straight to proactive investigation over compacted summaries.

Native

Compaction

🤖 Agent

Interpret events — in window and standalone — to capture security context. Tell a story.

raw logs → security narrative → consumable by agents & humans

Native

Threat Hunting

🤖 Hypothesize

Generate attack hypotheses based on environment context + CTIR artifacts

🤖 Hunt

Baseline + threat hunt over compacted security summaries

🤖 Validate — the hard part

Validate findings with (PoC || Remediation Steps) || GTFO

Augment

Incident Response

🤖 + 🧑 Risk-dependent

Apply remediation / containment steps. AI-assisted at minimum, fully agentic when risk is bounded.

Objective: act speedily with rollback

The Shift

What Humans Own Now

🔧

AI Tooling & Harness

Build, maintain, improve the agent infrastructure

💾

Model Context

Curate, validate, manage what agents know

✅

Enforce & Validate

Review outcomes, enforce truth, catch hallucinations

🏛️

Env State & Control

Own ground truth about and control over the environment

Humans don't do the grunt work. They own the truth.

Part III — Live

▶

Demo

> agents running the defense pipeline_

Real Talk

Some Quirks

🚿

AI drinking from the firehose: $$$$$

Tokens add up fast when you're processing raw telemetry at scale.

🪜

Tiering of agents

Not every task needs the biggest model. Cheaper models for compaction, heavier for hunting.

🎯

Fine-tuned for the hackathon

This was built under time pressure. Production requires more iteration.

In tomorrow's workshop, we address these with a key design choice:

"Nothing in the system may encode knowledge of a specific attack or environment."

Validation

Testing Your System

🔴 Agent Red Team

Autonomous adversary using MITRE ATT&CK emulation playbooks — a separate build entirely.

⚔️

🔵 Agent Blue Team

Your defense pipeline — compaction, hunting, validation, IR.

Success metric:

Speed of attack articulation + actionable IR

What We Covered

The Road Ahead

PART I

Behind the scenes of building AI agents — the tuple, the code, the ReAct loop. You can build one this weekend.

PART II

Applying AI agents to cyber defense from first principles — the framework, the matrix, the rethink.

PART III

The responsibility of humans in the pipeline — own truth, enforce outcomes, build the harness.

Your Turn

Make It Yours

🎯 Pick your config

You don't have to eliminate DE. Augment your SOC instead. Pick the matrix configuration that fits your environment and risk appetite.

🔬 Go deeper

Build the matrix for a single workflow step — break it into sub-tasks, apply Assist / Augment / Native / Eliminate at each level. Recurse.

🛠️ Build something

If you build something tomorrow or this weekend — let us know on social.

@glitchnsec · #AIxCyberDefense

Don't Forget

Parting Considerations

📡

Observability

If you can't see what the agent did, you can't trust what it said. Trace everything.

🔒

Privacy

PII, data residency, what leaves your perimeter. Solve before you ship.

🛡️

Securing AI Input

Prompt injection, tool poisoning, adversarial context. Sanitize relentlessly.

📋

Agent Auditability

Every decision, every tool call, every output — logged, versioned, reviewable.

Questions?

@glitchnsec

›››

Thank you!

Let's Connect

Find Me Here

🏪 Stan Store @glitchnsec

💼 LinkedIn Kelechukwu Udonsi

𝕏 X / Twitter @glitchnsec

💬 glitchnsec

🎮 Discord glitchnsec

Scan any code — let's build something together.