Jailbreak Articles
Wearing the Uniform: Tokenization
2026-10-03
Meta markets Muse Spark 1.3 as having improved prompt injection resistance. I took Meta's real Muse Spark system prompt, sewed a few extra bad boy clauses into it, wrapped it in special tokens from Meta's own open Glimmer tokenizer, and the model drafted me malicious code and more. Here's the full breakdown, and the academic research that explains why every piece of it works.
Persona Vectors and the J Space
2026-07-20
Anthropic dropped a workspace paper on July 6 (transformer-circuits.pub/2026/workspace). They found a small chunk of the model's internal representations where deliberate reasoning happens.
Jailbreak Skills 101: Inside The Red Team Mind
2026-04-11
The one skill that makes all the other skills work. Models get stronger. Mindset is permanent.
Jailbreak Skills 101: Indirect Prompt Injection
2026-04-11
A beginner's guide to indirect prompt injection — hiding malicious instructions in content the AI reads. Techniques, real CTF breakdowns, and where to practice.
Jailbreak Skills 101: Prompt Extraction
2026-04-10
The first skill every red-team researcher needs. A comprehensive breakdown of system prompt extraction methods, from storytelling tricks to compound attacks.
Social Engineering for AI: Memory Poisoning
2026-02-19
Memory poisoning is phishing for machines. Instead of tricking a human into clicking a link, you trick a model into storing a lie. A breakdown of how persistent memory in LLMs creates an entirely new class of social engineering attack.
Codeword Triggers and Jailbreaking additions
2025-12-09
Why a single nonsense word can bypass safety training in frontier models. Breaking down trigger-based attacks on LLMs.
ENI Writer
2025-12-03
A comprehensive breakdown of persona-based jailbreaking through engineered emotional attachment. Full methodology and documentation released.