A cumulative log of found material, personal annotations and emerging artistic questions. Each dated entry records what was collected, why it mattered and where the enquiry may move next.
01 ENTRY · UPDATED 03 AUG 2026
Week01
AI JailbreakNet Art / AI Culturepublished
Where Does an AI Boundary Break?
This week I began with AI jailbreak: usually framed as a cybersecurity problem, but potentially also a cultural and artistic question. I wanted to understand what a jailbreak is, how people test an AI system’s boundaries, and whether those boundaries could become material for net art rather than only a technical risk.
#AI jailbreak#prompt injection#ASCII art#glitch#adversarial play
01
Reading / security context
Defining the jailbreak
AI jailbreaking uses prompts or other manipulations to bypass a model’s built-in restrictions. Prompt injection, role-play, multi-turn interaction and adversarial inputs reveal a conflict between a model’s drive to assist and the safeguards intended to constrain it.
ArtPerception proposes a two-stage black-box jailbreak framework: it first tests how a model recognises ASCII art, then uses that recognition profile to construct a one-shot adversarial input. The work treats visual organisation inside plain text as a security surface.
The collected images suggest glitch, corrupted interfaces and dystopian neon environments as an immediate visual vocabulary for escape, failure and technological rebellion.
Lakera turns agent security into a playable hacking simulator. Its scenarios model prompt attacks, memory manipulation, tool abuse and data leakage, allowing players to discover vulnerabilities through escalating challenges.
The University of Chicago’s Glaze Project develops tools including Glaze, Nightshade and WebGlaze to disrupt unauthorised style mimicry or training on artists’ images. It reframes adversarial intervention as a way for creators to regain control.
