Rembg: Unauthenticated SAM model injection permits local-file disclosure through malicious ONNX graphs

The public `/api/remove` endpoint uses attacker-controlled `extras.sam_model` in both SAM filenames and GitHub download URLs. Path traversal plus a URL fragment lets an unauthenticated attacker place malicious ONNX models beside a target file. The models read the file as ONNX external data and encode its bytes into the returned PNG. This issue affects SAM's `sam_model` in 2.0.81 and returns complete file contents. <img width="2172" height="1182" alt="image" src="https://github.com/user-attachments/assets/70aacab1-5ef8-4aad-b30e-e3780adef4d2" /> *Screenshot 1 — First unauthenticated request primes the path and returns HTTP 500.* ### Technical details `rembg/commands/s_command.py:188-223,384-397` accepts public `extras` JSON and passes it to `new_session()`. `rembg/sessions/sam.py:313-355` inserts `sam_model` into encoder/decoder filenames and fixed GitHub URLs, then invokes `pooch.retrieve()` without a pinned hash. Requests removes the URL `#fragment`; the local POSIX path retains it. Pooch creates attacker-selected parent directories without enforcing cache containment. The first request creates the path tree and returns HTTP 500. The identical second request downloads the attacker asset twice and stores it as both SAM models beside the target. The graph uses a relative external tensor to read the target and maps every byte to mask pixels. rembg returns the mask as a PNG, which the attacker decodes back into the original bytes. <img width="2432" height="1254" alt="image" src="https://github.com/user-attachments/assets/99215c39-3619-4952-8f41-5ef2034f7499" /> *Screenshot 2 — The identical second request returns HTTP 200 with an encoded PNG.* <img width="2858" height="1548" alt="image" src="https://github.com/user-attachments/assets/26322071-101c-41aa-a356-08a4969ef475" /> ### Proof of concept The attached loopback-only PoC produced: ```text VULNERABLE: unauthenticated SAM model injection disclosed the synthetic /flag first-http=500 second-http=200 decoded-value=REMBG_AUDIT_SYNTHETIC_FLAG_2026_08_28 ``` The `/etc/passwd` variant produced: ```text VULNERABLE: disclosed /etc/passwd (839 bytes) root:x:0:0:root:/root:/bin/bash daemon:x:1:1:daemon:/usr/sbin:/usr/sbin/nologin ... nobody:x:65534:65534:nobody:/nonexistent:/usr/sbin/nologin ``` The decoded contents matched `/etc/passwd` inside the owned container. Testing used a local TLS stand-in for GitHub; no public target was contacted. <img width="1816" height="1146" alt="image" src="https://github.com/user-attachments/assets/664eaf25-5963-4ca6-b7eb-3e198499194a" /> *Screenshot 3 — The decoder reconstructs all 839 bytes of `/etc/passwd`.* ### Impact An unauthenticated attacker can execute arbitrary ONNX computation and disclose complete files that rembg can reach under the required filesystem permissions. The supplied root-running container exposed both `/flag` and `/etc/passwd`. ### Remediation Allow only fixed SAM identifiers mapped server-side to fixed filenames, URLs, and pinned hashes. Reject path and URL syntax; canonicalize destinations and enforce cache containment before creating directories. Run the container as a non-root user with a read-only root filesystem. [decode_samfragment_png.py](https://github.com/user-attachments/files/31601255/decode_samfragment_png.py) [samfragment_04.py](https://github.com/user-attachments/files/31601256/samfragment_04.py) [samfragment_etc_passwd_lab.py](https://github.com/user-attachments/files/31601258/samfragment_etc_passwd_lab.py)