deepstream-eval-and-finetune

This is a readable training and inference dashboard with same-origin API calls and local UI preference storage; it shows no clear malware, credential theft, or external exfiltration. Its main security concern is potential DOM-based XSS: API-derived fields, including dataset class names and QC or insight messages, are inserted into innerHTML without escaping, so attacker-controlled response data could execute in the UI origin. No external domains, IPs, or credential targets appear in the fragment. Severity is moderate and depends on whether untrusted users can influence the API data.

No direct malicious payload (e.g., exfiltration logic, credential theft, or a reverse shell) is shown in the fragment, but it has high-impact execution points: it conditionally executes an external setup script ($SK/app/setup.sh) and runs arbitrary command strings inside Docker via bash -c "$INNER". Without auditing setup.sh, the construction/control of $INNER, and the server code at /work/$SK/app/server:app, malicious or unsafe behavior cannot be ruled out. Treat as a meaningful supply-chain/execution-risk area rather than a benign configuration snippet.