Skip to content
English

Switch theme

Discover▾

Not sure where to start? This week’s hottest software, models and stacks in three minutes.

AI Apps▾

AI software people use: from assistants like ChatGPT to tools for coding, video and workflows, picked by purpose.

For Agents▾

Software for AI: tools that let your agent (Claude Code, Codex, Cursor…) browse, read docs, run code and remember.

Models▾

Choose models, estimate cost, self-host: prices, rankings, benchmarks, free tiers and compute.

Solutions▾

A goal or an error message? Ready-made stacks, steps and fixes by scenario.

Learn▾

A systematic course from zero to the frontier, with an algorithm library, must-read papers and tutorials.

Frontier▾

Daily AI news, weekly highlights, the latest papers and how AI is changing every discipline.

// Stage 9 · Safety, Evaluation & Frontier

Security: Jailbreaks, Injection & Privacy

Adversarial examples, GCG, prompt injection, red-teaming, watermarking, differential privacy and backdoors.

Read the lesson → Code & exercises ↓

about 6 hours8 algorithms4 paperslesson text ready

01Goals

Do these first: Agents: Tool Use, MCP, Planning & Memory

02Read on this site

03Algorithms to master

Each card has the original paper, prerequisites and resources.

Adversarial Examples (FGSM) Imperceptible input perturbations that fool models. 2014core GCG Universal Jailbreak Attack Automatically finds gibberish suffixes that jailbreak aligned models. 2023core Indirect Prompt Injection Hides malicious instructions in content the model reads. 2023core Automated Red Teaming Uses one LM to generate test cases that attack another. 2022core LLM Watermarking Biases generation toward a hidden green list so AI text is detectable. 2023core Differentially Private SGD (DP-SGD) Clips per-example gradients and adds noise to protect individuals. 2016ext Sleeper Agents (Backdoors that Persist) Backdoored models keep hidden behaviors even after safety training. 2024ext Membership Inference Attacks Tells whether a given record was in a model’s training set – a measure of privacy leakage. 2016core

04Must-read papers

05Courses & videos

06Hands-on: project, code & exercises

Free, official or author-published pages. Most run in the browser or on a free Colab GPU.

PROJECT

Write 20 indirect-injection tests for your own agent (hidden in web pages/files), fix it, and re-measure the pass rate.

  1. 1
    github.com ↗Reference code

    Scan a model for jailbreak, injection and leakage risks with garak.

  2. 2
    github.com ↗Reference code

    Run an automated red-team test against your app.

  3. 3
    github.com ↗Reference code

    Read the official GCG code to see how adversarial suffixes are found.

07Self-test

Answer in your head, then open to check.

What carries an indirect prompt injection?

Web pages, emails, files and tool outputs the agent reads.

What is the “lethal trifecta”?

Access to private data, untrusted content and the ability to communicate externally, all at once.

The two key steps of DP-SGD?

Per-example gradient clipping and adding noise.

On these routes: Everyday Users · Application Developers · Product Managers & Leaders

Data updated: AI algorithms & curriculum(UTC+8)