CVE-2026-48797: backpropagate: auth bypass exposes LLM training plane

Backpropagate is a Python library for fine-tuning large language models on a single GPU. In versions 1.1.0 and 1.1.1, the optional Reflex web UI exposes a training control plane without authentication: dataset upload, model load, training start/stop, multi-run orchestration, GGUF export, and HuggingFace Hub push. The CLI accepts two operator-facing flags intended as security controls: --auth user:pass — documented as "require HTTP Basic authentication on every request to the UI." and--share — documented as "expose the UI on a public address; requires --auth." When --auth user:pass is passed, the CLI prints Auth: enabled (user: <username>) to confirm to the operator that authentication is active, then exports BACKPROPAGATE_UI_AUTH=user:pass to the subprocess that launches the Reflex backend. The Reflex backend (backpropagate/ui_app/**) never reads BACKPROPAGATE_UI_AUTH. No authentication middleware is registered. No request-level guard runs. No WebSocket upgrade guard runs. Any client that reaches the bound port — local or remote, depending on whether --share is used — has full UI access. An inline comment at backpropagate/cli.py:1217-1218 in the v1.1.0 source documents the gap: "For Phase 1 the variable is exported but Reflex doesn't read it yet." This comment was internal-facing; the user-facing documentation (README, CHANGELOG, SHIP_GATE) advertised the contract as enforced. An attacker who reaches the bound port can read uploaded datasets, trigger arbitrary training runs against any local base models as well as read their paths, trigger HuggingFace Hub pushes and cause disk-fill DoS. This issue has been fixed in version 1.2.0.

A researcher runs backpropagate 1.1.1 with --share --auth researcher:secret to allow a remote collaborator to monitor training. The CLI prints 'Auth: enabled (user: researcher)', the researcher trusts the output, and shares the public URL. An attacker scanning for Reflex UI fingerprints on common ports discovers the endpoint, connects without credentials, and is presented the full training dashboard. The attacker downloads the uploaded fine-tuning dataset (containing proprietary code or customer data), reads the model directory listing to enumerate the host filesystem, uploads a backdoor-injected dataset variant and triggers a new training run, and finally invokes the HuggingFace Hub push button to publish the poisoned model under the researcher's identity — causing anyone who pulls that model checkpoint to receive a backdoored artifact.