AI provider Anthropic has released the system prompts of its Claude models. The company wants to create more transparency in AI development.
Anthropic has published the system prompts for its latest Claude AI models in a move to increase transparency in AI development. System prompts are typically considered trade secrets, and market leader OpenAI has not yet disclosed its own prompts. However, this is not a first for Anthropic, as Amanda Askell, the company's AI director, shared an earlier version of Claude 3's system prompt at X in March. Additionally, prompts are sometimes exposed through prompt injections or other attacks.
The recently published prompts, dated July 12, 2024, provide detailed information about the capabilities and limitations of the Claude 3.5 Sonnet, 3 Opus, and 3 Haiku models. They outline what the AI should and should not do, such as prohibiting facial recognition, opening links, and instructing the models to handle controversial topics objectively. The prompts also describe personality traits, including intellectual curiosity. Among the models, Claude 3.5 Sonnet has the most extensive prompt.
The prompts are accessible through the Claude apps and online. Anthropic intends to regularly publish updates to the prompts.
Anthropic wants to be more open than OpenAI
Founded in 2021 by former OpenAI employees, including siblings Dario and Daniela Amodei, who previously held high-level positions at OpenAI, Anthropic aims to be more transparent than OpenAI. Dario Amodei was a senior researcher at OpenAI for several years, leading the AI safety team for two years, while Daniela Amodei was responsible for security and policy at OpenAI from May to December 2020. Both were involved in developing OpenAI's GPT-3.
Recently, renowned AI security researcher Jan Leike also moved from OpenAI to Anthropic. Leike left OpenAI due to security concerns, criticizing the company's security culture.
Anthropic places a strong emphasis on AI safety and recently announced an expanded bug bounty program focused on identifying and defending against universal jailbreak attacks. The company offers rewards of up to $15,000 for discovering new security vulnerabilities.
