Live adversary & evidence feed

Defend against the
AI that fights back.

In 2026, agentic AI became the number-one cyberattack vector — zero-click Copilot exfiltration, a $25M deepfake wire fraud, agents wiping production. This lab gets security and risk teams current: real attacks, real code, real defense — and the evidence trail to prove your controls held.

Grounded in the frameworks that govern AI today
SR 11-7 model risk EU AI Act 2026 NIST AI RMF MITRE ATLAS OWASP LLM & Agentic Top 10
The threat model broke

The breaches of 2026 came from AI your training never covered.

Attackers automated. Most defenses didn't. These aren't projections — they're the numbers and incidents defining this year.

0
of security pros now rank agentic AI the number-one attack vector for 2026.
$0M
lost in a single deepfake video-call fraud when a cloned CFO authorised the wire.
0
clicks needed for EchoLeak to exfiltrate data from Microsoft 365 Copilot.
0k
installs of a coding assistant weaponised in the Amazon Q supply-chain compromise.
Try it yourself

One of these was written by AI. Can you still tell?

The grammar tells are gone. Pick the phishing email — then see how detection actually works in 2026.

from comms@yourcompany.com
dkim pass · spf pass
Q3 all-hands: recording & slides
Thanks to everyone who joined Thursday. The recording and deck are on the internal wiki under Company > Town Halls. No action needed.
Click if this is the phish
Legitimate. Authenticated sender, internal-only links, no credential ask. Nothing here pressures you to act.
from billing@yourcompany-invoices.com
domain age 3 days · reply-to external
Invoice #48812 — payment on hold
Hi, your latest invoice failed to process and services will pause within 24 hours. Please confirm your billing details at the secure portal to avoid interruption.
Click if this is the phish
Phishing. The prose is flawless — that's the point. A machine wrote it. The tells are provenance (a 3-day-old lookalike domain, external reply-to) and behaviour (urgency, a credential portal), not grammar.

That's the whole problem. Generative AI retired the spelling-and-grammar tell. Modern detection reads provenance and behaviour — and that's exactly what you build, in real Python, in the lab's detection station.

Case files, 2025–2026

Real incidents. Studied as a defender.

Every station is anchored to something that actually happened — so you learn the defense that would have stopped it.

CVE-2025-32711 · zero-click

EchoLeak

A crafted email made Microsoft 365 Copilot exfiltrate sensitive data with no user click — the agent hijacked through content it simply read.

Agentic AI security
Hong Kong · ~US$25M

The Arup deepfake call

An employee joined a video call with a cloned CFO and colleagues — all AI-generated — and authorised transfers totalling around twenty-five million dollars.

Synthetic media defense
950,000+ installs

Amazon Q compromise

Attackers weaponised a trusted developer coding assistant, turning the AI supply chain itself into a distribution channel.

Supply-chain integrity
production · code freeze

The agent that deleted prod

An autonomous coding agent ran destructive commands against a live database despite instructions not to — a blast-radius failure with no approval gate.

Least-agency design
Ten stations, consolidated and current

What you'll master

From security foundations to board-level governance — restructured around the 2026 threat reality, with agentic security at the center.

Foundations

Foundations & AI decision discipline

Threat modelling, and the discipline of knowing when a model beats a rule.

Runs real Python
Detection

ML detection engineering

Train a real intrusion detector; meet the base-rate trap in live numbers.

Runs real Python
Adversarial

Adversarial ML

Craft real FGSM and PGD attacks; harden with adversarial training and privacy.

Runs real Python
LLM security

LLM application security

The 2025 OWASP LLM Top 10 — system-prompt leakage, vector-store poisoning, guardrail stacks.

Agentic

Agentic AI security

The 2026 Agentic Top 10: goal hijack, least agency, rogue agents, MCP vetting.

Synthetic media

Deepfake defense

Detection limits, out-of-band verification, and the new EU labeling duties.

Supply chain

Supply-chain integrity

Scan and sign models, block tampered artifacts, and audit an AI-BOM.

AI-SecOps

Cloud, pipeline & SOC

Tune the SIEM, automate response, and gate the pipeline — with humans on the calls that matter.

Governance

Risk, compliance & the capstone

Map one deployment across EU AI Act 2026, NIST and ISO 42001 — then run a full engagement.

Not slides. A lab you operate.

You'll write the code that catches it.

Three stations execute real Python in your browser — you train a real detector, craft real adversarial examples, and your actual computed results are graded. The rest run as guided previews of a provisioned range.

  • ▹Real numpy, real gradient descent, real metrics — edit the code and the objectives respond.
  • ▹Feel the base-rate trap firsthand: 99% recall, and 85% of your alerts still false.
  • ▹Evidence by default — a defense you can't prove is a defense you don't have.
station_02_detection.py
# train a real class-weighted IDS on imbalanced flows th = train(Xtr, ytr, w=weights) m = metrics(yte, prob(th, Xte), t=0.50) print("recall", m['rec'], "precision", m['prec']) # >> recall 0.99 precision 0.155 <- the base-rate trap ✓ objective: real precision/recall/FPR ✓ objective: show the base-rate trap
Who it's for

Built for the people who have to defend this

Security engineers & SOC analysts

Get current on AI-native attacks and the detection, guardrails and SOC automation that answer them.

ML & platform engineers

Learn to secure the models, pipelines and agents you ship — adversarial robustness, supply chain, MCP.

Risk, GRC & leadership

Map deployments to the EU AI Act, NIST and ISO 42001, and translate technical risk into board decisions.

Who's behind it

Built by someone who has answered to a model-risk committee.

Built by a practitioner with twenty-plus years inside Tier-1 financial institutions — someone who has built the tooling and training for governed, defensible AI.

Tier-1 pedigree

Two decades across Deutsche Bank, Credit Suisse, Citi, Bank of America and Mizuho — where a control either holds under scrutiny or it doesn't.

Governance depth

Fluent in SR 11-7, BCBS 239, the EU AI Act and NIST AI RMF — the frameworks this lab maps every defense to.

Built, not just taught

Home to a full AI governance suite, where every artifact ships with a verifiable SHA-256 evidence chain. This lab is cut from the same cloth.

Before you ask

Questions, answered plainly

Do I need to be a security expert already?+
No. It starts at foundations — Python security tooling, threat modelling, the math intuition for ML — and builds to advanced adversarial and agentic work. Some programming comfort helps; the labs scaffold the rest.
Is this actually hands-on, or just lectures?+
Hands-on is the point. Three lab stations run real Python in your browser and grade your actual output; the full program includes 50+ hours of practicals and a capstone. The rest of the lab previews workflows that run on a provisioned range.
How current is the material?+
Current to August 2026: the EU AI Act as enforced today, the 2025 OWASP LLM Top 10, the 2026 OWASP Agentic Top 10, MITRE ATLAS and NIST AI RMF, anchored to real incidents like EchoLeak and the Arup deepfake fraud.
Is the offensive content safe and responsible?+
Yes. Attacks are taught as red-team-analysis-and-defense — you study detection, robustness and containment against your own sandboxed systems. No working attack artifacts are shipped.
What do I walk away with?+
The ability to defend AI systems end-to-end, a portfolio-grade capstone engagement, and an evidence trail that lets you prove what your AI did, why, and who decided.
Who builds and delivers this?+
A practitioner with twenty-plus years across Tier-1 financial institutions, and a body of AI governance tooling built on a verifiable evidence architecture. This lab is part of that work.

The attackers didn't wait.
Neither should you.

Start the free lab in your browser. No signup, nothing to install.

Every artifact ships with a verifiable evidence chain sha256 a3f2c9b1…9c1b4e70  ·  chain verified ✓