User extracts hidden system prompts by spawning subagents
Aryvyo · x · 2026-09-09
A user found that by getting a model to spawn subagents through some indirect route, the otherwise-hidden system prompts can be extracted — and noted how dense they are. Another reminder that prompt secrecy in current products is largely security theater.
Related event: Users Find Way to Bypass Codex's Encrypted Subagent Prompts(2 posts)→
More from Fun
- GPT-6 Astra turns 9 crappy studio photos into a fully interactive 3D model, demo stuns researchers — CSProfKGD · 2026-09-09
- Even plumbing companies are adding "preferred source" buttons now — gaganghotra_ · 2026-09-09
- A simulated fly brain can play Beat Saber: RL post-training of biological brains in sim — ccerrato147 · 2026-09-09
- SfM session so packed at ECCV 2026 that organizers opened a second room with a second mic — ducha_aiki · 2026-09-09
- Gemini describes writing as "weighting millions of semantic pathways," not inspiration — mimi10v3 · 2026-09-09
- Submitting your paper to an AI review tool and it finds mistakes everywhere — Faheem_uh · 2026-09-09