CVE-2026-48797

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. If developers cannot immediately upgrade to 1.2.0 run backprop ui with no flags so it binds to localhost, use SSH port-forwarding (ssh -L 7860:localhost:7860 <training-host>) instead of --share for remote access, and audit any host previously launched with --share, re-issuing any HF tokens used during those sessions.

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. If developers cannot immediately upgrade to 1.2.0 run backprop ui with no flags so it binds to localhost, use SSH port-forwarding (ssh -L 7860:localhost:7860 <training-host>) instead of --share for remote access, and audit any host previously launched with --share, re-issuing any HF tokens used during those sessions.

Summary

In backpropagate >= 1.1.0, the optional Reflex web UI (pip install backpropagate[ui], launched via backprop ui) exposes a training control plane: 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."
• --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.

This advisory is filed primarily because the runtime contradicted an operator-facing security claim. Code-only bugs of comparable shape (auth check missing entirely from a path) would already warrant disclosure; the additional false-promise dimension raises the severity.

Impact

An attacker who reaches the bound port can:

• Read uploaded datasets rendered in the UI preview, including content of any JSONL/CSV/TXT file the legitimate operator has uploaded for fine-tuning.
• Trigger arbitrary training runs against any base model the operator has installed locally or that can be downloaded from HuggingFace.
• Trigger HuggingFace Hub pushes to repositories named via the UI input (subject to the operator's local HF token's scope — typically all repos owned by the operator).
• Cause disk-fill DoS via the rx.upload endpoint (no size cap, no extension filter, no per-session count cap in v1.1.0 / v1.1.1).
• Read model paths (source_model_path, dataset_path, model, uploaded_path) which are user-supplied and bypass the safe_path() helper that lives in backpropagate/ui_security.py (path validation is dead code on the Reflex surface in v1.1.0 / v1.1.1).

The combination of unauthenticated training control, HF push target spoofing, and path-input traversal makes the affected endpoint suitable for both data exfiltration (reading uploaded training data) and supply-chain attacks (pushing tampered model weights to the operator's HF account).

The local-only default (no --share) reduces exposure to a host-local attacker. The --share flag is documented as a "public URL" feature; operators who used --share --auth user:pass had no warning that the auth half was inert.

Patches

Fixed in v1.2.0 (released 2026-05-23). The patch implements real ASGI middleware via rx.App(api_transformer=basic_auth_transformer) that gates HTTP routes AND the /_event WebSocket upgrade. Four modes (no_auth_local_only / token_auto / explicit_creds / production), HMAC-signed cookie validated PRE-websocket.accept(), Host + Origin allowlists. The middleware ships alongside a 4-layer defense in depth at the cli.py / ui_app/app.py / rxconfig.py / env-strip surfaces so direct python -m reflex run invocations (bypassing the CLI guard) also enforce authentication.

Upgrade with:

• pip: pip install --upgrade backpropagate
• npm: npm install -g @mcptoolshop/backpropagate@latest

Full release notes: https://github.com/mcp-tool-shop-org/backpropagate/blob/main/CHANGELOG.md#120---2026-05-23

Workarounds

If users cannot upgrade…

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. If developers cannot immediately upgrade to 1.2.0 run backprop ui with no flags so it binds to localhost, use SSH port-forwarding (ssh -L 7860:localhost:7860 <training-host>) instead of --share for remote access, and audit any host previously launched with --share, re-issuing any HF tokens used during those sessions.

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. If developers cannot immediately upgrade to 1.2.0 run backprop ui with no flags so it binds to localhost, use SSH port-forwarding (ssh -L 7860:localhost:7860 <training-host>) instead of --share for remote access, and audit any host previously launched with --share, re-issuing any HF tokens used during those sessions.

Summary

In backpropagate >= 1.1.0, the optional Reflex web UI (pip install backpropagate[ui], launched via backprop ui) exposes a training control plane: 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."
• --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.

This advisory is filed primarily because the runtime contradicted an operator-facing security claim. Code-only bugs of comparable shape (auth check missing entirely from a path) would already warrant disclosure; the additional false-promise dimension raises the severity.

Impact

An attacker who reaches the bound port can:

• Read uploaded datasets rendered in the UI preview, including content of any JSONL/CSV/TXT file the legitimate operator has uploaded for fine-tuning.
• Trigger arbitrary training runs against any base model the operator has installed locally or that can be downloaded from HuggingFace.
• Trigger HuggingFace Hub pushes to repositories named via the UI input (subject to the operator's local HF token's scope — typically all repos owned by the operator).
• Cause disk-fill DoS via the rx.upload endpoint (no size cap, no extension filter, no per-session count cap in v1.1.0 / v1.1.1).
• Read model paths (source_model_path, dataset_path, model, uploaded_path) which are user-supplied and bypass the safe_path() helper that lives in backpropagate/ui_security.py (path validation is dead code on the Reflex surface in v1.1.0 / v1.1.1).

The combination of unauthenticated training control, HF push target spoofing, and path-input traversal makes the affected endpoint suitable for both data exfiltration (reading uploaded training data) and supply-chain attacks (pushing tampered model weights to the operator's HF account).

The local-only default (no --share) reduces exposure to a host-local attacker. The --share flag is documented as a "public URL" feature; operators who used --share --auth user:pass had no warning that the auth half was inert.

Patches

Fixed in v1.2.0 (released 2026-05-23). The patch implements real ASGI middleware via rx.App(api_transformer=basic_auth_transformer) that gates HTTP routes AND the /_event WebSocket upgrade. Four modes (no_auth_local_only / token_auto / explicit_creds / production), HMAC-signed cookie validated PRE-websocket.accept(), Host + Origin allowlists. The middleware ships alongside a 4-layer defense in depth at the cli.py / ui_app/app.py / rxconfig.py / env-strip surfaces so direct python -m reflex run invocations (bypassing the CLI guard) also enforce authentication.

Upgrade with:

• pip: pip install --upgrade backpropagate
• npm: npm install -g @mcptoolshop/backpropagate@latest

Full release notes: https://github.com/mcp-tool-shop-org/backpropagate/blob/main/CHANGELOG.md#120---2026-05-23

Workarounds

If users cannot upgrade…

Summary

In backpropagate >= 1.1.0, the optional Reflex web UI (pip install backpropagate[ui], launched via backprop ui) exposes a training control plane: 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."
• --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.

This advisory is filed primarily because the runtime contradicted an operator-facing security claim. Code-only bugs of comparable shape (auth check missing entirely from a path) would already warrant disclosure; the additional false-promise dimension raises the severity.

Impact

An attacker who reaches the bound port can:

• Read uploaded datasets rendered in the UI preview, including content of any JSONL/CSV/TXT file the legitimate operator has uploaded for fine-tuning.
• Trigger arbitrary training runs against any base model the operator has installed locally or that can be downloaded from HuggingFace.
• Trigger HuggingFace Hub pushes to repositories named via the UI input (subject to the operator's local HF token's scope — typically all repos owned by the operator).
• Cause disk-fill DoS via the rx.upload endpoint (no size cap, no extension filter, no per-session count cap in v1.1.0 / v1.1.1).
• Read model paths (source_model_path, dataset_path, model, uploaded_path) which are user-supplied and bypass the safe_path() helper that lives in backpropagate/ui_security.py (path validation is dead code on the Reflex surface in v1.1.0 / v1.1.1).

The combination of unauthenticated training control, HF push target spoofing, and path-input traversal makes the affected endpoint suitable for both data exfiltration (reading uploaded training data) and supply-chain attacks (pushing tampered model weights to the operator's HF account).

The local-only default (no --share) reduces exposure to a host-local attacker. The --share flag is documented as a "public URL" feature; operators who used --share --auth user:pass had no warning that the auth half was inert.

Patches

Fixed in v1.2.0 (released 2026-05-23). The patch implements real ASGI middleware via rx.App(api_transformer=basic_auth_transformer) that gates HTTP routes AND the /_event WebSocket upgrade. Four modes (no_auth_local_only / token_auto / explicit_creds / production), HMAC-signed cookie validated PRE-websocket.accept(), Host + Origin allowlists. The middleware ships alongside a 4-layer defense in depth at the cli.py / ui_app/app.py / rxconfig.py / env-strip surfaces so direct python -m reflex run invocations (bypassing the CLI guard) also enforce authentication.

Upgrade with:

• pip: pip install --upgrade backpropagate
• npm: npm install -g @mcptoolshop/backpropagate@latest

Full release notes: https://github.com/mcp-tool-shop-org/backpropagate/blob/main/CHANGELOG.md#120---2026-05-23

Workarounds

If users cannot upgrade…