A full system prompt of the Claude 3.7 Sonnet has leaked

In early May 2025, the full system prompt of Anthropic’s Claude 3.7 Sonnet model was made publicly available. This document of about 24,000 tokens gives unique access to the internal architecture of one of the most advanced AI assistants on the market.

What leaked

The Claude 3.7 Sonnet system prompt is not just a set of instructions. It includes:

• Detailed behavioral directives such as striving for neutrality, avoiding categorical judgments, and using Markdown for code formatting.
• Filtering mechanisms and XML tags to structure responses and ensure security.
• Instructions for using tools, including web search, artifact generation, and interacting with external APIs.
• Protocols to protect against “jailbreaks” and unwanted behavior.

This prompt is ten times the size of previously published versions and is essentially an operating system for Claude, defining his behavior in various scenarios.

Why it matters

A leak of this magnitude raises questions about security and transparency in AI development:

• Security: If a model’s internal instructions can be revealed, it jeopardizes protection against manipulation and attacks.
• Transparency: On the one hand, details about the model’s operation can contribute to user trust. On the other hand, revealing such details can be exploited by attackers.
• Ethics: Understanding how the AI makes decisions is important to assess its objectivity and lack of bias.

Community Reaction

Following the leak, there was a lot of discussion in the AI development community. Many expressed concern that such leaks could be a source of new vulnerabilities. Others believe that this is a chance to improve security practices and increase transparency in AI development.

What’s next

Anthropic has previously stated a commitment to “constitutional AI”, aiming to create models that focus on security and ethics. However, the current leak highlights the need to rethink approaches to protecting the inner workings of AI.
With the increasing integration of AI into various spheres of life, ensuring security and transparency becomes a priority. Developers will have to find a balance between openness and intellectual property protection.

For those who want to read the full leaked prompt, it’s available on GitHub.