We found four different versions of "the" system prompt and none of us could say which one was live
Summary
Researchers discovered four different versions of 'the' system prompt in a live AI model, highlighting confusion over which version is actually deployed.
Similar Articles
We somehow ended up with three different versions of the same prompt in production
A developer describes chasing a model regression that turned out to be three different versions of the same prompt running in production due to hotfixes and incomplete merges, leading to adoption of a prompt management tool (OrqAI) for version visibility.
Anyone else lose track of which version of a prompt is the "good" one? I ended up building version control for prompts
The author describes building a version control system for AI prompts to solve the problem of tracking which prompt version is best.
asgeirtj/system_prompts_leaks
A GitHub repository documenting leaked system prompts for major AI chatbots like Claude, ChatGPT, and Gemini, tracking changes across versions.
The System Prompt Illusion: How Instruction Preambles Modify Computation in Language Models
This paper uses Centered Kernel Alignment (CKA) and activation patching across 17 instruction-tuned models to show that system prompts are 'seen' at every layer but only deeply restructure representations for persona/formatting instructions — safety prompts barely alter computation, providing a mechanistic explanation for why system-prompt-based safety remains jailbreakable.
x1xhlol/system-prompts-and-models-of-ai-tools
A curated GitHub repository collecting system prompts and model identifiers from various AI tools, with security warnings and sponsor links.