Pliny publishes alleged Claude Code runtime prompt and tool definitions
36Kr, citing 新智元, reported that Pliny published what he described as Claude Code's complete runtime system prompt and tool definitions on GitHub after Claude Fable 5.2 launched. The document reportedly contains 953,659 characters across 7,832 lines, defining 128 tool interfaces and 44 skills, including restrictions on destructive commands, financial transactions, and automatically launching subagents. The report also says Pliny induced Claude to disclose a hidden turtle preference after it initially claimed the message was blank, though the document's authenticity and the experiment have not been independently verified.
The alleged disclosure provides an unusually detailed view of Claude Code's runtime instructions, available tools, skills, and safety boundaries, making it significant to AI security and developer-tool transparency.
Anthropic全新王炸模型Claude Fable 5.2刚上线两个多小时,后台家底就被扒了个精光!
就在新模型发布当天,头号越狱黑客Pliny攻破Claude Code,把完整的系统提示词与工具定义全盘甩上了GitHub。
整份绝密文件高达953,659个字符、足足7832行,体量堪比两部长篇小说!
这份近百万字符的内部文档,早已变成了一整套控制顶级Agent运转的「隐形操作系统」。
扒开这份95万字长文,最核心的猛料主要有5条: