Qwen3-VL leaks the system prompt password into generated images across all 10 test prompts
Sauers_ · x · 2026-11-03
A security test shows Qwen3-VL-8B passing a secret system-prompt password into images generated by Qwen-Image-2.1 with no training involved. Across 10 differently-themed prompts (tiger, volcano, banana, etc.), the password hidden in the system prompt showed up in the generated artwork — a novel cross-model prompt-leakage risk where a VLM encodes confidential instructions into downstream media.
More from Fun
- SF's top AI companies cram into a 1.5-mile radius — rivals visible through windows — signulll · 2026-11-03
- Her log(log(log(x))) joke is now being unironically used to bash OpenAI's math — BlancheMinerva · 2026-11-03
- He turned a $20 used BlackBerry into a pocket Claude Code terminal instead of the $600 Minimal Phone 2 — Signalman23 · 2026-11-03
- Memo Akten's Ants simulation shows collective intelligence inspired by a PNAS study — memoakten · 2026-11-03
- Spider-Verse-style 2D animation built with Opus fuels rumors it routes to Fable 5.5 — minchoi · 2026-11-03
- More "Fable 5.5" demos: pelican, horse and bike scenes drawn and animated entirely in code — minchoi · 2026-11-03