Please report security vulnerabilities privately rather than through public GitHub issues, discussions, or pull requests.

• GitHub private reporting: If enabled for this repository, use Report a vulnerability.
• Email: Contact aixu2008nice@gmail.com with CNS Rover SECURITY in the subject line if private reporting is unavailable or you prefer email.

Include the affected commit or version, operating system and relevant dependency versions, the impact, and minimal reproduction steps. For a problematic scenario, model, or recording, describe its source and provide the smallest necessary example. Redact credentials and private information from logs and attachments.

Reports will be reviewed to assess impact and coordinate a fix and disclosure where appropriate. Please allow time for investigation before publishing exploit details. Response times are not guaranteed. Fixes target the current default branch; when practical, check whether the issue also affects that revision.

CNS Rover consists of a Python simulator, a local FastAPI HTTP/WebSocket server, and a React browser workbench. It reads scenario and neural-model files and persists recordings, camera frames, telemetry, and a SQLite run index on disk. Optional data-preparation workflows download connectome assets or query neuPrint through the vendored R package.

Reports of particular interest include:

• Cross-site scripting or injection through scenario names, telemetry, recordings, or other content rendered in the workbench.
• Bypasses of HTTP/WebSocket host and origin checks, including browser requests that expose local recordings or trigger unauthorized simulation actions.
• Path traversal or unintended filesystem access through recording, replay, export, model-loading, or archive-extraction paths.
• Unsafe parsing of JSON, NPZ, Feather, CSV, or downloaded archives, including code execution and practical resource-exhaustion attacks.
• Failures in download integrity checks, exposure of neuPrint credentials, or dependency and supply-chain vulnerabilities with a credible project impact.

The provided serve command binds to 127.0.0.1. The workbench has no user accounts, authentication, or tenant isolation. Host and origin checks provide some protection against browser-origin attacks; they are not authentication. Keep the service local. Public hosting, shared-machine access, or reverse-proxy exposure requires a separate access-control and deployment review.

Custom controllers loaded with --controller module:Class execute Python code with the server process's permissions. The optional R exporter also executes local R code. These extension mechanisms are not sandboxes; use trusted code. Load models, scenario files, and recordings from trusted sources. Checksums check files against recorded or pinned values; a manifest supplied alongside an untrusted model does not establish its publisher's identity.

Browser recordings are stored under runs/workbench/ by default, and CLI runs use the selected output directory. Recordings may contain task text, scenario settings, camera frames, and controller debug output. Review them before sharing. Keep neuPrint tokens in the environment or the upstream client's credential configuration, and out of committed files and reports. Running the bundled native model does not require a token.

Controller collisions, timeouts, and navigation failures without a security impact are ordinary bugs or research limitations and can be reported in public issues. Include the scenario, controller, model identity, and relevant recording.

CNS Rover currently provides simulation only. Its obstacle checks and simulated watchdog do not constitute a physical emergency stop or a real-robot safety system. Inference runs synchronously, so the simulator cannot interrupt a controller that hangs during a decision. Hardware integration requires separate process isolation, deadlines, actuator safeguards, and safety validation.