What I liked about this Claude Code Sandbox Escape we reported to Anthropic was that it required chaining together multiple vulnerabilities: some "standard," and others requiring the harness to call a specific tool to automatically trigger the next link in the chain, using