AI Red-Teaming Academy

Ein Selbstlernpfad durch Prompt Hacking, Jailbreak-Techniken und modellseitige Abwehrmaßnahmen, geschrieben während der Teilnahme an Gray Swans Hazard Hunt und HackAPrompt.

Grundlagen

Modul 1Grundlagen7 min read
A Taxonomy of Prompt Hacking
Prompt injection, jailbreaking and prompt leaking get used interchangeably online, but they target different things and call for different defenses. This module draws the boundaries.
PROMPT-INJECTIONJAILBREAKINGTAXONOMYLLM-SECURITY
Read

Fortgeschritten

Modul 2Fortgeschritten8 min read
The Jailbreak Technique Library
Roleplay framing, many-shot context stuffing, encoding tricks, logic traps and context poisoning: five distinct ways to make a model's safety training miss the request in front of it.
JAILBREAKINGROLEPLAY-FRAMINGMANY-SHOTCONTEXT-POISONING
Read