Defending AI Agents
עברית Join waitlist

Free practical course · self-paced · in preparation

DefendingAI agents

Build agents you know how to defend. Learn to control tool actions, test defences and investigate failures, with code you build yourself. For developers, ML engineers and data practitioners who want to take responsibility for agent security. No security background needed.

Formatself-pacedModules22CoursefreeStatusin preparation

Join the waitlist View curriculum

01What you will learn

Turn your ML skills into security decisions.

You already work with models, data and uncertainty. Learn to apply that experience to a new set of questions: which actions need protection, which alerts deserve attention and what evidence is enough to trust a defence.

01

Decide which alerts matter

Use base rates to work out how many alerts are real attacks, and how much investigation your team can handle.

02

Choose a workable threshold

Balance attacks caught against legitimate work blocked. Set thresholds against an explicit false-alarm budget.

03

Test on the data agents use

Evaluate on documents, tool calls, JSON and logs. Find out what changes when production data differs from your test set.

04

Evaluate a policy before rollout

Replay recorded traffic and test in shadow mode to assess a policy before it starts blocking actions.

02Try the reasoning

Catching attacks is only half the problem.

Suppose a detector catches 99% of attacks and flags 1% of harmless calls. Is that useful? Change the attack frequency and call volume to see what those numbers mean for the team reviewing the alerts.

Share of alerts that are real attacks
1%

Alerts per day
Real attacks detected
Real attackFalse alarm
03What you will build

A working security layer for an AI agent.

Start with a small agent and build the controls around it: detection, enforcement, evaluation and incident response. Use open, general-purpose tools so you can inspect the code, change it and explain each decision.

  1. M01A tool-calling agent and an attack you can reproduce
  2. M03A gateway that checks tool calls before they run
  3. M04Prompt-injection detectors using patterns and a trained classifier
  4. M08An evaluation harness to measure attacks caught and legitimate tasks blocked
  5. M11Controls that track untrusted data and restrict what it can influence
  6. M12Scoped credentials, a sandbox and a kill switch
  7. M16A signed audit log and replay tools to investigate decisions
  8. M18An incident response runbook tested against a recorded attack
04From lab to production

A small error rate can become a full-time job.

Build and measure locally, then calculate what your results imply at higher volumes. Learn to account for review time, false alarms and capacity before deciding whether a control is practical to operate.

An illustrative workload
101alerts from 10,000 calls

With one attack per 10,000 calls, a detector that catches 99% of attacks and flags 1% of harmless calls produces about 101 alerts. Only about one corresponds to a real attack.

At 50 million calls a day
500,000false alarms per day, approximately

At the same attack frequency and false-alarm rate, reviewing each false alert for 30 seconds would take about 520 eight-hour workdays, every day. Use this calculation to set an operational budget.

05The curriculum

From your first attack to incident response.

22 modules and a capstone, about 248 hours in total. Study at your own pace in English or Hebrew. Each module includes practical work and measurements you can explain. The course is in preparation; join the waitlist for the first releases.

I

Foundations

  1. 00Orientation: security decisions under uncertainty
  2. 01Set up your lab and build a small agent
  3. 02Security fundamentals for ML and data practitioners
  4. 03Agent architecture and attack surfaces
II

Attacks

  1. 04Prompt injection and jailbreaks
  2. 05Poisoning, backdoors and agent memory
  3. 06Privacy attacks and differential privacy
III

Measurement

  1. 07Detection: base rates, false-alarm budgets and thresholds
  2. 08Reliable evaluation and distribution shift
  3. 09Adversarial testing and adaptive attackers
  4. 10Detecting anomalies in behaviour and action sequences
IV

Enforcement

  1. 11Designing boundaries for untrusted input
  2. 12Identity, least privilege and containment
  3. 13Policy decisions and human approval
  4. 14Supply chain security: models, packages and tools
  5. 15Personal data: detection, redaction and erasure
V

Operations

  1. 16Audit trails, replay and policy evaluation
  2. 17Learning policy from traffic safely
  3. 18Monitoring and incident response
  4. 19Governance, audit and regulation for engineers
  5. 20Agent discovery, identity linkage and coverage
  6. 21Operating controls: latency, capacity and failure
CapstonePut your design to the test.Build a gateway for a new scenario and test it against attacks you have not seen. Explain how it would operate at scale, then write an adversarial test report on another learner’s work.
06Access and assessment

Learn for free. Add a certificate if you choose.

All lessons, labs and automated checks will be free. The optional final assessment and certificate are paid, with personal review and written feedback.

Free

The full courseNo account needed to read.

  • Every lesson, lab and automated check
  • English and Hebrew
  • A starter kit for each module
  • Materials you can keep and revisit
Paid

Final assessment and certificateOptional. Price announced at launch.

  • Personal review of your capstone
  • Detailed written feedback
  • A signed certificate that others can verify
  • An optional public page linking to your reviewed work
07Before you join

Frequently asked questions.

Who is this course for?

Developers, ML and applied AI engineers, data scientists and analysts who connect models to data and tools. It focuses on building and evaluating controls, from a tool call to incident response.

Do I need a security background?

No. You need Python and the basics of statistics or machine learning. Security concepts are taught from the ground up, and a free four-hour preparation module covers web fundamentals.

What equipment do I need?

A laptop with at least 8 GB of memory running macOS, Linux or Windows with WSL 2. No GPU is required. A dedicated module walks you through Docker and lab setup.

Does it teach a particular security product?

You build the components with open, general-purpose tools. The lessons explain the choices and name alternatives, so you can understand the approach and adapt it to your own stack.

What is free, and what is paid?

All lessons, labs and automated checks will be free. Only the optional final assessment and certificate are paid. You can complete the learning material without paying for assessment.

What languages are available?

English and Hebrew. Code, commands and tool names stay in English in both versions.

When can I start?

The course is in preparation. Join the waitlist to hear when the first modules are published. There is no launch date to announce yet.

Can my team get started sooner?

The live workshop is available separately: a four-hour session for up to 15 people, at your office or on Zoom. It covers an attack, defences and evaluation, and does not require the course.

Hear when the
course opens

You are on the list. We will write when the first modules are out.

That does not look like an e-mail address. Check it and try again.

Too many attempts from your network. Please try again in an hour.

The list could not be reached. Please try again, or write to shimon@defendingaiagents.com.

08  Waitlist

Join the waitlist for an email when the first modules are published, plus occasional progress updates. Your address is used only for this course and is never shared. To leave the list, email shimon@defendingaiagents.com.