System Prompts Are Not Secrets
Hidden instructions to deployed models will keep being extracted and published shortly after release, pushing builders to treat the prompt layer as public and move protection to server-side controls and model training.
weakening · confidence 38 · Emerging (watchlist) · tracking since November 21, 2026 · updated November 22, 2026
Score history
Daily conviction score, 0 to 100. Higher means the thesis is more strongly corroborated.
Now 38 · -2 since Nov 21 · ranged 38 to 40
Why the conviction moved
- Nov 21Context
OpenAI's claimed drop in reasoning-leak rates on GPT-6.2 Astra, from 54% to 0.2% under adaptive extraction attacks, comes from training plus server-side account enforcement. That is the shift of protection to the model and backend that the thesis predicts. It covers reasoning traces, though, not system prompts, which keep leaking.
- Nov 21Strengthened +3
The public GitHub repository asgeirtj/system_prompts_leaks published extracted system prompts for Claude Fable 5.2 alongside ChatGPT and Gemini, down to behavioral rules such as a ban on cursing unless the user asks or swears heavily. A newly released flagship's prompt layer was exposed shortly after launch, as the thesis predicts.
Source trail
Supporting · November 21, 2026
A public GitHub repository publishes extracted system prompts for Claude Fable 5.2 and other production models
The public GitHub repository asgeirtj/system_prompts_leaks published extracted system prompts for Claude Fable 5.2 alongside ChatGPT and Gemini, down to behavioral rules such as a ban on cursing unless the user asks or swears heavily. A newly released flagship's prompt layer was exposed shortly after launch, as the thesis predicts.
AI Post (Telegram)Context · November 21, 2026
OpenAI says GPT-6.2 Astra leaks protected reasoning in 0.2% of adaptive extraction attacks, down from 54% on GPT-6 Astra
OpenAI's claimed drop in reasoning-leak rates on GPT-6.2 Astra, from 54% to 0.2% under adaptive extraction attacks, comes from training plus server-side account enforcement. That is the shift of protection to the model and backend that the thesis predicts. It covers reasoning traces, though, not system prompts, which keep leaking.
OpenAI News
Unlock full source trail, score history, and daily updates.
Unlock TrendsTownsquare
Argue the thesis in Townsquare.