tensorlake

This module is best characterized as an evaluation harness with a significant sandbox/trust-boundary weakness rather than intrinsic malware. It runs an external `claude` tool with `--dangerously-skip-permissions` and then reads local files based on paths extracted from untrusted tool-use outputs, serializing their contents into `files.json`. If the subprocess/tool output is influenced maliciously, the harness can enable unintended local file disclosure and sensitive-data retention. No direct malicious payload is evident within this Python file itself, but the integration materially increases security risk.