Prompt
injection
Untrusted content becomes an instruction.
Security / AI / Reality
01
Cybersecurity Is Dead
Exploit got cheaper, faster, and ....
The human in the loop
02
Who
I am
Builder / Breaker
The asymmetry
03
AI makes security unequal.
Attackers pick one target.Defenders cover everything.
Published CVE records / calendar year
7.5× in 10 years
Economics of repetition
04
One automated runApprox.
$3 - 4
Less than a slice of cake.
Not every exploit. Not every model. But cheap inference makes failed attempts disposable.
The disclosure gap
05
The exploit can arrive before the advisory.
FixA maintainer patches quietly.
DiffA bot watches every public commit.
InferThe changed path reveals the bug.
ExploitPoC now. Advisory later.
Trust, inherited
06
We are built on top of a supply chain...
Your appthe visible layer
Frameworkstrusted by default
Packagesupdated constantly
Models + dataopaque artifacts
Unknown maintainerone token away
New system. New failure modes.
07
AI created a new attack surface.
Prompt
Rogue
agents
A model with tools turns confusion into action.
Backdoored
models
Weights and serializers become executable supply chain.
08
Closed
A few companies can gate access, training, telemetry, and policy.
Open
Everyone can inspect and build - but provenance, poisoned data, and model integrity become your problem.
Looks equal. Isn't.
The talent paradox
09
Engineers have never mattered more.
AI replaces amplifies engineers.
AI lowers the floor.
Engineering judgment raises the ceiling.
Live case study
10
Build →
Break.
One short demo.
Build the feature.
Find the assumption.
Turn it into impact.
CodeAttackLesson
A practical ending
11
What do we do?
Assume the attacker automates first.
Verify every supply-chain artifact.
Give agents the minimum possible power.
Test models like production code.
Invest in engineers who understand systems.
Make attacks expensive again.