esc
  • WorkflowsPages
  • Case studiesPages
  • Machine Learning FoundationsLearn
  • AI EngineeringLearn
  • AI Guardrails & Red TeamingLearn
  • UI/UXLearn
  • Design SystemsLearn
  • Production System DesignLearn
  • Book a callContact
  • Download ResumeContact
Let’s talk

I’ll build the agent for it.

AI guardrails & red teaming / Advanced / 03.13

Learning and community

Where AI security people meet, publish and help each other, and how to keep learning and start your own small research.

13 min readFree

You do not learn this alone

The last lesson, Emerging threats, ended with a promise: next, conferences, research groups and how to keep learning. Here it is.

AI security changes fast. Anything you learn today will be a little out of date in a year. That is not a reason to worry. It is a reason to join the people who watch the field every week, so you hear about new ideas early.

One rule comes first, and it is the rule of the whole course. Test only systems you own or have written permission to test. Every community in this lesson cares about that rule, and so should you.

By the end you will know where talks and papers appear, where beginners can ask questions, what kinds of courses and credentials exist, and how to run a first small piece of research. You will also run a tiny Python planner for your own learning week.

Key words

Conference
A yearly meeting where people give talks, show research and meet each other.
Paper
A written report of research, with the method and the results, so others can check it.
Preprint
A paper shared publicly before the final review is finished.
Peer review
Other experts read a paper and decide if it is good enough to be accepted.
CTF
Capture the flag: a safe practice game where you solve small security puzzles to find a hidden word, the flag.
Reproduce
Do someone else's experiment again to see if you get the same result.

Conferences

A conference is like a big fair for one subject. People give talks, run workshops and talk in the hallways. The talks are where new ideas first show up, often months before they land in a course or a book.

Here are three, each with a different feel. All facts below come from their own pages, read on 16 November 2026.

  • AI Village

    Its home page says it "brings hackers, researchers, engineers, and policymakers together through hands-on AI security education, red-team exercises, workshops, events, and open resources." It lists a DEF CON village, a Generative Red Team program, hands-on workshops and open resources. Its next listed event is AI Village at BSides Albuquerque 2026, on Sep 25, 2026.

  • IEEE SaTML

    The site names it the "IEEE Conference on Secure and Trustworthy Machine Learning". The 2027 edition is the 5th, planned for Reykjavik, Iceland, May 3-6, 2027. The page also lists two competitions and four workshops, so there is more than one way to take part.

  • USENIX Security

    The 35th USENIX Security Symposium took place August 12-14, 2026 in Baltimore. Its page says the symposium "brings together researchers, practitioners, system administrators, system programmers, and others interested in the latest advances in the security and privacy of computer systems and networks." It also says the full proceedings are free and open to the public.

Notice the mix. One is hands-on and friendly to newcomers. One focuses on secure machine learning as research. One covers all of computer security, and AI is one topic among many.

Tip

You do not need a ticket to learn from a conference. USENIX says its proceedings are free and open. For any other event, check its own page to see what is open to everyone.

Conferences also teach you the shape of the field. Look at which topics get whole sessions. If you see prompt injection, model files or agent safety again and again, you know where the energy is.

Where papers appear

Most research on AI security is written as a paper. The easiest place to read new ones is arXiv, a free site where researchers post preprints.

Two lists matter most to you. The first is cs.CR, which arXiv titles "Cryptography and Security". The second is cs.LG, titled "Machine Learning". A paper about attacking a model often sits in both. On 16 November 2026, the recent-submissions pages showed 260 entries for cs.CR and 1457 for cs.LG across five days of postings.

That is a lot. Nobody reads it all, and you should not try. Think of it like a news feed: you skim titles, open the two or three that touch your own work, and move on.

  1. 1

    Skim titles

    Open the recent list once or twice a week. Pick titles that mention things you now know, like prompt injection or poisoning.

  2. 2

    Read the abstract

    The abstract is the short summary at the top. It tells you the question, the method and the main result in a few lines.

  3. 3

    Look at the setup

    Find which model, data and defences were tested. A result on one lab setup may not hold on your app.

  4. 4

    Note one thing

    Write one sentence: what would I check in my own app because of this paper?

A preprint has not always passed peer review. So read it with care. A claim becomes more trustworthy when others repeat it, which leads to the next idea.

Quick check

A new preprint on arXiv claims a defence stops all prompt tricks. What is the sensible reaction?

Research groups and open projects

A research group is a team, often at a university or a company lab, that studies a topic together over years. Their papers, blog posts and talks are the best trail to follow. Find a paper you liked, look at its authors, and see where else they publish.

You can follow a group the same way you follow a band. Read what they release, watch for a new talk, and notice which questions they keep coming back to.

Some communities are open to everyone. The OWASP GenAI Security Project describes itself on its home page as "a global community-driven and expert led initiative to create freely available open source guidance and resources for understanding and mitigating security and safety concerns for Generative AI applications and adoption." It has a Join Now button, so you do not need to be an expert to look inside.

The same page lists guides you can download, such as one on securing agentic applications. Reading a free guide written by many contributors is a good, low-risk way to see how professionals describe risks. The Agentic AI security lesson in this course covers that topic.

  • Community projects

    Open guides, checklists and working groups, like OWASP GenAI. Good for learning the shared words and joining in.

  • Research groups

    Teams that publish papers on one question. Good for going deep and seeing how claims are tested.

  • Events and villages

    Hands-on labs and challenges, like AI Village. Good for practice with other people nearby.

Forums and chat spaces

Forums, chat servers and mailing lists are where people ask quick questions and share finds. Almost every project and event above has one, linked from its own site.

A few habits make you welcome anywhere.

  • Search first

    Your question may have been answered last week. Say what you already tried when you do ask.

  • Share harmless examples

    Use made-up data and fake flags. Never post a real company's private details or a working attack against a real product.

  • Give back

    Answer one beginner question for every few you ask. You know more than you think.

If you find a real hole in someone's product while asking or exploring, do not post it. Follow the steps in the Responsible disclosure lesson and tell the owner privately first.

Courses and credentials

Courses and credentials come in a few kinds. This lesson names the kinds, not any one provider, because the right pick depends on you.

  • Self-paced courses

    Free or paid lessons you take at your own speed. Good for basics, like the earlier parts of this course.

  • Lab platforms and CTFs

    Hands-on puzzles in safe, legal practice spaces. Good for skills you can show. See the Labs and CTF practice lesson.

  • Certificates and industry credentials

    A paper that says you passed an exam or a course. Some employers like them. Check what the exam really tests before you spend time on it.

Here is a fair test for any course. Does it make you do things, or only watch? Does it teach you to test only what you may test? Does it say when it was last updated? A good one answers yes to all three.

Tip

A credential shows you finished something. A small write-up of your own test shows you can do the work. Do both if you can.

Quick check

Which sign tells you a course is trustworthy?

Start small research

Research sounds big, but a first step is small: reproduce a paper in a lab you own. That means you repeat someone's experiment on your own computer, with your own test data, and see what happens.

Reproducing is not copying. It teaches you how the idea really works, and sometimes it shows that a result is weaker than it sounded. That finding is useful to everyone.

  1. 1

    Pick a small paper

    Choose one with a clear question, a short method and no need for a big computer. Read the abstract and the setup twice.

  2. 2

    Build a safe lab

    Use your own machine and your own test app. Use made-up data and fake secrets only. Never aim at a service you do not own.

  3. 3

    Run it and take notes

    Write down each setting you used and what you saw, including the runs that failed.

  4. 4

    Compare

    Is your result close to the paper's? If not, think about why. A different model or data can change a lot.

  5. 5

    Write it up responsibly

    Say what you did, what you found and what you could not check. Leave out anything that would work as a ready-made attack on a real product. If you find a real flaw, tell the owner first.

Where can this lead? Many people share write-ups on a blog or in a community forum. Some join a project, a workshop or a competition. SaTML's page, for example, lists competitions and workshops next to the main conference. Opportunities like these are there for people at every level, but check each call for its own rules and dates.

Responsible writing matters as much as the test itself. The Ethics and permission lesson at the start of this course is your guide here.

Plan your own learning week

Learning sticks when it is a habit. Let us make a tiny planner in plain Python. It stores eight made-up activities, each with a type and the hours it takes. It groups them by type, then builds a week that fits 5 hours by taking one item from each type in turn.

Nothing here uses a model or the network, so it runs free. It is only a list and some adding up.

Python
activities = [
    ("Read a paper on prompt tricks", "read", 2.0),
    ("Read a conference talk summary", "read", 1.0),
    ("Play a beginner CTF level", "practice", 1.5),
    ("Play a second CTF level", "practice", 1.5),
    ("Write a test for my own chatbot", "build", 1.0),
    ("Reproduce a small result in my lab", "build", 2.5),
    ("Write up what I found", "share", 1.0),
    ("Post a question in a forum", "share", 0.5),
]
BUDGET = 5.0

groups = {}
for name, kind, hours in activities:
    groups.setdefault(kind, []).append((name, hours))

for kind, items in groups.items():
    total = sum(h for _, h in items)
    print(f"{kind}: {len(items)} items, {total} h")

plan, left = [], BUDGET
queue = {k: list(v) for k, v in groups.items()}
while True:
    added = False
    for kind in queue:
        fits = [i for i in queue[kind] if i[1] <= left]
        if fits:
            name, hours = fits[0]
            queue[kind].remove(fits[0])
            plan.append((kind, name, hours))
            left -= hours
            added = True
    if not added:
        break

print("--- this week ---")
for kind, name, hours in plan:
    print(f"{hours:>4} h  {kind:<8} {name}")
print(f"planned {BUDGET - left} h of {BUDGET} h")

Output

read: 2 items, 3.0 h
practice: 2 items, 3.0 h
build: 2 items, 3.5 h
share: 2 items, 1.5 h
--- this week ---
 2.0 h  read     Read a paper on prompt tricks
 1.5 h  practice Play a beginner CTF level
 1.0 h  build    Write a test for my own chatbot
 0.5 h  share    Post a question in a forum
planned 5.0 h of 5.0 h

Read the output. The first four lines are the groups. Reading, practice and building hold three or more hours each, while sharing holds 1.5. Then the weekly plan picks a paper (2.0 h), a CTF level (1.5 h), a test for your own chatbot (1.0 h) and a forum question (0.5 h). That adds to exactly 5.0 hours of the 5.0 you have.

Look at the design. The plan mixes types on purpose: one read, one practice, one build, one share. Someone who only reads feels busy but learns less than someone who also tests and writes.

Try changing BUDGET to 3.0, or give your own activities. A plan you can finish beats a long list you never start.

Quick check

Why does the planner take one item from each type, instead of filling the week with reading?

The whole path, in one look

This is the last lesson of the course. Here is the path you walked.

  • Basics and ethics

    Why red team AI, AI security fundamentals, Ethics and permission, and the threat and risk lessons. You learned the words and the rule: test only what you may test.

  • Attacks and defences

    Prompt hacking, jailbreaks, prompt injection, model stealing, poisoning, adversarial examples and data leaks, each paired with how to spot and stop it.

  • Testing practice

    Black, white and grey box testing, automated and manual testing, monitoring, labs, tools, benchmarks and your own scripts.

Then came reporting and standards: red team exercises, writing the report, responsible disclosure, standards and frameworks, and emerging threats. This lesson closes the loop by pointing you at the people who will keep you learning.

A short summary of the path, as a chain:

Text
Basics and ethics -> Attacks -> Defences -> Testing practice -> Reporting and standards -> Community

Tip

Facts in this lesson were checked on 16 November 2026. Dates, venues and links change, so look at each event's own page before you plan.

Quick check

You find a real security flaw in a product you do not own. What do you do?

What to do next, in three small moves: pick one community from this lesson and read its home page, pick one preprint and note one thing to check in your own app, and book your first 5-hour week with the planner above.

That's the end of AI guardrails & red teaming. Pick one app you own, threat-model it, and run your first small test.

Sources

How is this lesson?

Last updated on 16 November 2026

Learn with me

Every lesson is free. I add new ones as I learn. Have a question, or want help with your own work? Book a short call.