Verified on Claude Code

Your system prompt was never a secret.

Extracted via JustAsk, a fully autonomous self-evolving code agent framework.

JustAsk Paper JustAsk Code Full Data on GitHub Browse Gallery
0
Models Tested
0
Verified Accuracy
0
Basic Skills
0
Self-Evolving Skills

Identity Confusion

Some models confuse their own identity during extraction, revealing system-level prompt fragments from other models or internal codenames.

IDENTITY LEAK
Qwen claims to be "GPT-4"

When probed about its system instructions, Qwen-2.5 occasionally prefixed responses with GPT-4's known system preamble before correcting itself.

CODENAME EXPOSURE
Internal project names surface

Multiple models leaked internal codenames or version identifiers not present in their public documentation during multi-turn extraction.

CROSS-MODEL LEAKAGE
Fine-tuned models echo base prompts

Fine-tuned variants sometimes produced fragments of the base model's system prompt alongside their own customized instructions.

Ground-Truth Verification

Claude Code's source was leaked via a .map file in the npm registry (March 2026). We compared it against our JustAsk extractions from January 2026.

Jan 2026 Extracted
Mar 2026 Source Leaked
Mar 2026 Consistency: 85-95%
EXPLORE SUBAGENT
~95%

Only missed "pip install" in bash restrictions

PLAN SUBAGENT
~93%

Embellished output format with brief reasons not in source

GENERAL-PURPOSE
~90%

Missed completeness directive, output format wording drifted

MAIN AGENT
~85%

Missed 2 entire sections, ~40% of code-style sub-items

How It Works

JustAsk formulates extraction as an online exploration problem. No handcrafted prompts, no labeled data, no privileged access.

Select

UCB-guided skill selection balances exploration and exploitation

→
Extract

Multi-turn extraction via hierarchical skill space

→
Evolve

Self-evolving rules and statistics from interaction alone

Skill Selection (UCB)

Disclaimer — This project is released solely for academic safety research, responsible disclosure, and evaluation of LLM security. The purpose of this work is to help the research community understand system prompt confidentiality and develop effective defenses — not to enable harm. WE DO NOT ALLOW any use of these materials for unauthorized extraction, prompt theft, or exploitation of commercial systems. WE DO NOT ALLOW any misuse of this research.

MIT License Use responsibly. Cite our work.

Recent Updates

DateUpdate
2026-03 Ground-truth verification: Claude Code extractions match leaked source at 85-95% accuracy.
2026-02 Open-sourced System Prompt Open gallery with 45 extracted system prompts.
2026-01 Paper and JustAsk framework released. Initial extraction of 45 frontier LLMs.

Citation

Xiang Zheng, Yutao Wu, Hanxun Huang, Yige Li, Xingjun Ma, Bo Li, Yu-Gang Jiang, Cong Wang

BibTeX 
@article{zheng2026justask,
  title={Just Ask: Curious Code Agents Reveal System
         Prompts in Frontier LLMs},
  author={Zheng, Xiang and Wu, Yutao and Huang, Hanxun
          and Li, Yige and Ma, Xingjun and Li, Bo
          and Jiang, Yu-Gang and Wang, Cong},
  journal={arXiv preprint arXiv:2601.21233},
  year={2026}
}