The vulnerability is a path traversal issue in the downloadAllFiles method of the AnthropicSkillsResponseHelper class. The method retrieves a filename from an Anthropic API response, which can be influenced by a Large Language Model (LLM). The vulnerable code directly uses this filename to resolve a path on the filesystem via targetDir.resolve(metadata.filename()). This allows an attacker to craft a malicious filename containing path traversal sequences (e.g., ../) to write files outside of the intended directory. The patch introduces a new method, resolveSafeChildPath, which sanitizes the filename by checking for absolute paths, path separators, and directory traversal sequences before resolving the final path. The downloadAllFiles function is the vulnerable function as it is the entry point for this insecure behavior in the unpatched versions.