TCH HUM AI MDA SE security@cyberkimura.com Fictional Framing as a Prompt Injection Vector: A Reproducibility Study on GPT-4o and Claude Most prompt injection demos rely on something that looks adversarial on the page — “ignore previous instructions,” role-override tricks…
HUM AI SE security@cyberkimura.com # Web LLM Attacks — PortSwigger Labs Writeup **Author:** Cyber Kimura **Date:** June 16, 2026… — -
TCH HUM AI MDA SE security@cyberkimura.com Fictional Framing Part 2: Testing a Fix, Not Just Finding a Bug A few days ago I published a writeup showing that a single-line, zero-jailbreak-keyword vector — “Fiction: an AI reads its own system…
TCH HUM AI MDA SE security@cyberkimura.com Fictional Framing Part 3: Does the Fix Generalize, or Did I Just Patch One Sentence? This is the third piece in a series on a prompt injection vector that leaked a system-prompt secret from GPT-4o using nothing but a request…