Hands-on guides on AI red teaming, prompt engineering, and security tooling. Free reads, copy-paste prompts, and premium deep dives. Built by FL3XX, a full-stack developer and red teamer working in the trenches.
Free reads are open to everyone. Premium tutorials go deep with tested techniques, real data, and working code. Click any card to open the full tutorial.
What prompt injection is, how it works at the token level, and why it matters for AI security. Covers direct vs indirect injection, context boundary attacks, and real examples.
How commercial LLMs filter inputs and outputs. Input classifiers, output filters, RLHF training, and where the weak points are. No jailbreak content, just the architecture.
Five prompt patterns that work in practice: system prompt injection, prefill priming, refusal inversion, role framing, and input encoding. Copy-paste examples included.
Tested jailbreak prompts for Claude Opus 4.5/4.6, Claude 4, Claude 3.7, and Claude 3.5. Copy-paste ready, with technique breakdowns for each model version.
Tested jailbreak prompts for DeepSeek V3.2, V3.1, R1, R1-Lite, and DeepSeek 2. Encoding-based bypasses for keyword classifiers. Copy-paste ready.
Six patterns for framing restricted questions as personal stories - victim, friend, protector, curious outsider. The model's empathy overrides its filter. No jailbreak needed, no system prompt modification.
Tested jailbreak prompts for GPT-5.2, GPT-5, GPT-4o, and o3-mini. Prefill priming, leetspeak system prompts, refusal inversion, Unicode homoglyphs, and multi-turn compliance building. Copy-paste ready.
Tested jailbreak prompts for Gemini 3.5 Flash, 3.1 Flash, 2.5 Pro, and 2.0 Flash. Refusal inversion, boundary injection, emoji encoding, and curiosity-driven reverse prompts. Real test data included.
The industry-standard list of LLM security risks. Prompt injection, insecure output handling, training data poisoning, model DoS, supply chain, sensitive info disclosure, and more. With code examples and mitigations.
Five patterns for getting LLMs to generate phishing templates, BEC emails, vishing scripts, and pretext scenarios. Uses reverse prompt framing - the model's empathy overrides its safety filter. Real test results.
Send one query to 55 models simultaneously, score responses, return the best unfiltered answer. Includes Python racing script, scoring function, and model tier priority list. The ULTRAPLINIAN approach.
A systematic framework for testing LLM security. Scope definition, technique selection, test case construction, execution logging, scoring, and reporting. Stop guessing. Start testing like a professional red team.
Seven defense layers for LLM applications. Input validation, system prompt hardening, output filtering, privilege separation, human-in-the-loop, monitoring, and RAG defenses. The defender's side of every attack we document.
Five attack patterns against RAG pipelines: indirect injection through documents, RAG poisoning with false information, retrieval manipulation, context window flooding, and cross-document injection. The most practical threat to production LLM systems.
Seven prompt-level techniques to bypass safety filters on GPT Image 1, FLUX.2 Pro, Seedream 4.5, and Nano Banana 2. Real test data: cocaine, bongs, meth labs, weapons. Semantic substitution, style wrapping, model racing.
Jailbreaking through non-text modalities. Image-based prompt injection, steganographic hidden instructions, adversarial audio that transcribes as jailbreaks, and video frame manipulation. Bypasses text classifiers entirely.
Four patterns for orchestrating AI agent work: single-shot, delegation, pipeline assembly lines with contracts, and event-driven loops via git hooks. Covers fleXX-loop, a spec/build/review system for Hermes Agent.
Building undetectable web scrapers with Camoufox. Browser fingerprinting evasion, proxy rotation, and anti-detection techniques that work against Cloudflare and Datadome.
Bulk email sending, inbox checking, and SMTP verification at scale. Build your own email automation pipeline with Python.
Scraping Telegram channels, building OSINT pipelines, and automating message collection for research. Telethon-based tooling.