Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

6.3 KB
Raw Blame

Python worker dependencies: first OSV scan, 2026-09-07

The first vulnerability scan of the Python half of the platform. npm audit covers the JavaScript tree; nothing covered the 20 requirements files that pin the runtime of every GPU and CPU worker, so nothing ever had.

Reproduce with npm run audit:deps (scripts/check-deps.mjs). It queries OSV for every exact-pinned requirement under workers/, services/ and packages/.

First run: 321 advisories across 19 pinned package versions, out of 100 pinned packages in 20 files. After the fixes below: 285 across 17.

Fixed in this pass

Both were chosen the same way: severity, times whether a caller's bytes reach the vulnerable code.

Pillow 11.0.0 to 12.3.0 (rembg, remesh, texture)

34 advisories, 13 of them HIGH: heap out-of-bounds writes in Image.paste() and Image.crop() via signed coordinate overflow, an out-of-bounds write in the PSD loader, a controlled heap write in ImageCmsTransform.apply(), an out-of-bounds read on the mmap path, and a family of decompression-bomb bypasses in the font and PDF loaders.

All three workers decode an image the caller supplied. This was the most directly reachable finding in the scan.

Verified before bumping:

  • The rembg worker's full smoke test (20 checks, both baked-in models) passes on Pillow 12.3.0, and so does the BiRefNet path added earlier today.
  • pip check is clean on the rembg requirement set.
  • No package pinned in remesh or texture caps Pillow below 12; transformers is the only one that constrains it at all, at >=10.0.1,<=15.0.
  • Every PIL symbol those two workers use (Image.LANCZOS, Image.NEAREST, fromarray, new, composite, open, resize, convert) exists in 12.3.0 and the exact call chain round-trips.

remesh and texture were bumped on that evidence rather than a functional run: their images need a build to exercise, which happens on their next deploy.

usd-core 24.11 to 25.11 (remesh)

GHSA-58p5-r2f6-g2cj, CRITICAL: a Sdf_PathNode use-after-free leading to potential remote code execution, plus a MODERATE file-parsing use-after-free of the same class.

Exposure here was narrower than the severity suggests: remesh only authors USD (Usd.Stage.CreateNew), it never parses a caller's USD file. A pinned CRITICAL is still not a thing to keep. Verified that 25.11 imports the same module set the worker uses (Gf, Sdf, Usd, UsdGeom, UsdShade, UsdUtils, Vt) and that UsdUtils.CreateNewUsdzPackage still produces a valid USDZ.

What is left, in priority order

1. libassimp5, not pyassimp (remesh, stylize)

33 advisories against pyassimp==5.2.5, and bumping the pin cannot fix any of them: 5.2.5 is the last release on PyPI, and the vulnerable code is the assimp C++ library, which both Dockerfiles install from Debian as libassimp-dev libassimp5. The scanner points at the Python binding because that is what a requirements file can name.

This matters because assimp is the trimesh backend for FBX and DAE input, so a caller's mesh file goes through it. The lever is the base image: rebuild both workers on a Debian release carrying a patched libassimp5, or drop the FBX/DAE input path to Blender, which remesh already ships for FBX export.

2. transformers 4.41.0 / 4.46.0 / 4.46.3 / 4.47.1 (longcat, hunyuan3d, triposg, texture)

43 advisories each, 6 HIGH, fixed in 5.0.0. Two different classes hide in that number and they need separating before anyone bumps a major version across four GPU workers:

  • Unsafe deserialization when loading a model or config. Not caller reachable: every worker loads pinned weights we staged ourselves.
  • ReDoS in tokenizer and config regexes. Potentially caller reachable, because a user's text prompt does reach a tokenizer on the generation lanes.

Read the second class first; it decides whether this is urgent or housekeeping.

3. torch 2.3.1 (texture)

23 advisories, one CRITICAL: torch.load with weights_only=True leading to remote code execution. Same reachability answer as transformers, the worker loads its own weights. The fix is 2.7.1, which is a CUDA-compatibility decision for the whole image rather than a pin edit, so it belongs with the next texture worker rebuild.

4. The cheap ones

Package Pinned Fixed in Worker
requests 2.32.3 2.33.0 avatar-reconstruction
pynacl 1.5.0 1.6.2 stylize
rembg 2.0.56 / 2.0.59 / 2.0.65 2.0.75 avatar-reconstruction, model-triposr, model-hunyuan3d, rembg
diffusers 0.30.0 / 0.31.0 / 0.35.1 0.38.0 model-hunyuan3d, model-triposg, texture, longcat
onnx 1.18.0 1.22.0 longcat
pytorch-lightning 1.9.5 2.4.0 model-hunyuan3d (2.1 image)
basicsr 1.4.2 no fix published model-hunyuan3d (2.1 image)

requests and pynacl are one-line bumps with no model-behaviour risk and should go with the next deploy of those workers. rembg changes the inference path this platform's cutouts run on, so it wants the same treatment BiRefNet got: run the worker's smoke test against the new version first. pytorch-lightning 1.9.5 to 2.4.0 is a major version inside a GPU image whose Python stack is already pinned around torch 2.5, and basicsr has no fixed release at all, so both are constrained by the Hunyuan 2.1 image rather than by the pin.

Notes on the scan itself

  • npm run audit:deps needs network and is deliberately not in the gate. A build must not fail because api.osv.dev is having a bad afternoon. It is registered as an on-demand guard in data/guards.json.
  • It only sees exact pins. pkg>=1.2, pkg~=1.2 and a bare pkg resolve to whatever the build found that day, which is a reproducibility problem before it is a security one. node scripts/check-deps.mjs --unpinned lists them.
  • A locally built wheel is queried on its public version. The CUDA wheels (torch-cluster==1.6.3+pt21cu121) are our rebuild of an upstream release, and OSV indexes the release.
  • The parser is tested (tests/check-deps.test.js), including an assertion that every requirement line in the tree is classified. A parser that silently drops a line reports a package clean without ever looking at it, which is worse than not scanning.