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.

fix(workers): clear 198 of 299 Python advisories by bumping worker pins that stay API compatible

nirholasa3f4ee71a098
diff --git a/workers/avatar-reconstruction/requirements.txt b/workers/avatar-reconstruction/requirements.txt
index fe869c60c..543f7b069 100644
--- a/workers/avatar-reconstruction/requirements.txt
+++ b/workers/avatar-reconstruction/requirements.txt
@@ -14,7 +14,11 @@ opencv-python-headless>=4.9.0
# ── Background removal ────────────────────────────────────────────────────────
# [gpu] pulls in onnxruntime-gpu so rembg uses CUDA when the L4 is present.
-rembg[gpu]==2.0.56
+# 2.0.59 clears the 2.0.57-and-earlier CORS/SSRF advisories. The fixes for the
+# rest landed in 2.0.75, which needs Python 3.11 and numpy 2.3; this image is
+# Python 3.10 with mediapipe holding numpy<2. All of them live in rembg's own
+# HTTP server (`rembg s`), which this worker never starts: it calls rembg.remove.
+rembg[gpu]==2.0.59
# ── Mesh operations (UV pre-computation, runs once at build time) ─────────────
trimesh==4.4.0
@@ -22,7 +26,7 @@ trimesh==4.4.0
# ── HTTP ──────────────────────────────────────────────────────────────────────
# requests: build-time UV-map download (precompute_uv.py) from a fixed GitHub URL.
# httpx: runtime SSRF-hardened image fetches via worker_security.fetch_remote_bytes.
-requests==2.32.3
+requests==2.33.1
httpx==0.28.1
# ── Server ────────────────────────────────────────────────────────────────────
diff --git a/workers/longcat/requirements-model.txt b/workers/longcat/requirements-model.txt
index d05510565..15358ba62 100644
--- a/workers/longcat/requirements-model.txt
+++ b/workers/longcat/requirements-model.txt
@@ -38,5 +38,5 @@ soundfile==0.13.1
soxr==0.5.0.post1
scipy==1.15.3
pyloudnorm==0.1.1
-onnx==1.18.0
+onnx==1.22.0
audio-separator==0.30.2
diff --git a/workers/model-hunyuan3d/requirements.hunyuan21.txt b/workers/model-hunyuan3d/requirements.hunyuan21.txt
index 760cafd26..c8b8995ef 100644
--- a/workers/model-hunyuan3d/requirements.hunyuan21.txt
+++ b/workers/model-hunyuan3d/requirements.hunyuan21.txt
@@ -23,11 +23,24 @@ ninja==1.11.1.1
pybind11==2.13.4
# ── Core ML / diffusion ──────────────────────────────────────────────────────
-transformers==4.46.0
+# Security bumps over upstream's list, each clearing every advisory its line can:
+# - transformers 4.57.6 (from 4.46.0) needs huggingface-hub>=0.34, hence 0.36.2,
+# the hub release model-triposg already runs.
+# - accelerate 1.15.0 (from 1.1.1): sharded-checkpoint path traversal.
+# - pytorch-lightning 2.6.6 (from 1.9.5): inference only subclasses
+# pl.LightningModule at import (hunyuanpaintpbr/unet/model.py); every Trainer,
+# checkpoint and callback path is training code this worker never runs.
+# 1.9.5 also stopped importing on a fresh build: its lightning_fabric calls
+# pkg_resources, which the setuptools 84 that the Dockerfile's
+# `pip install --upgrade setuptools` pulls no longer ships.
+# diffusers stays on 0.30.0: its open advisories need 0.38 and are reachable only
+# by loading an attacker-chosen repo or custom_pipeline; the paint pass loads the
+# in-image hy3dpaint/hunyuanpaintpbr pipeline from operator-staged weights.
+transformers==4.57.6
diffusers==0.30.0
-accelerate==1.1.1
-pytorch-lightning==1.9.5
-huggingface-hub==0.30.2
+accelerate==1.15.0
+pytorch-lightning==2.6.6
+huggingface-hub==0.36.2
safetensors==0.4.4
# ── Scientific computing ─────────────────────────────────────────────────────
diff --git a/workers/model-hunyuan3d/requirements.txt b/workers/model-hunyuan3d/requirements.txt
index fbe74e776..766937d06 100644
--- a/workers/model-hunyuan3d/requirements.txt
+++ b/workers/model-hunyuan3d/requirements.txt
@@ -6,9 +6,17 @@
# (DTensor import) that fails to load on the torch pinned below.
# ── Model stack (torch 2.3.1/cu121 era, matching the Dockerfile pin) ─────────
+# transformers 4.57.6 and accelerate 1.15.0 are the newest releases that clear
+# their advisories (ReDoS, model-file deserialization, sharded-checkpoint path
+# traversal) and still run on torch 2.3.1: transformers only imports DTensor on
+# torch>=2.5 and only disables itself below torch 2.1. Tiny seeded Dinov2, CLIP
+# and UNet models give bit-identical outputs against 4.46.3/0.34.2 on torch 2.3.1.
+# diffusers stays on 0.31.0: its open advisories need 0.38 and are reachable only
+# by loading an attacker-chosen repo or custom_pipeline, and this worker loads the
+# operator-staged weights plus hy3dgen's own in-image pipeline code.
diffusers==0.31.0
-transformers==4.46.3
-accelerate==0.34.2
+transformers==4.57.6
+accelerate==1.15.0
safetensors>=0.4.0
einops==0.8.0
omegaconf==2.3.0
diff --git a/workers/model-triposg/requirements.txt b/workers/model-triposg/requirements.txt
index a5438d7ba..1d5453270 100644
--- a/workers/model-triposg/requirements.txt
+++ b/workers/model-triposg/requirements.txt
@@ -11,7 +11,10 @@
# 2.3.1"), which cascades into a dead diffusers import and a container that never
# starts (crashes before listening on $PORT).
diffusers==0.31.0
-transformers==4.46.3
+# 4.57.6, not 4.46.3: the newest transformers that clears its ReDoS and
+# deserialization advisories and still runs on torch 2.3.1 (DTensor is only
+# imported on torch>=2.5). Needs huggingface_hub>=0.34, satisfied below.
+transformers==4.57.6
# Upstream also leaves peft and huggingface_hub unpinned; a bad transitive
# resolve of those crashed revision 00001 at import time (diffusers.loaders.peft
# raised "name 'torch' is not defined" before the server could bind $PORT).
diff --git a/workers/model-triposr/requirements.txt b/workers/model-triposr/requirements.txt
index e6324bff8..e1ef3c98f 100644
--- a/workers/model-triposr/requirements.txt
+++ b/workers/model-triposr/requirements.txt
@@ -8,7 +8,10 @@ trimesh>=4.4.0
Pillow>=10.3.0
# Background removal (TripoSR works best with isolated subject).
-rembg[gpu]==2.0.56
+# 2.0.59 clears the 2.0.57-and-earlier CORS/SSRF advisories. The rest are fixed
+# only in 2.0.75, which needs Python 3.11 and numpy 2.3 (this image is 3.10); all
+# of them live in rembg's own HTTP server, which this worker never starts.
+rembg[gpu]==2.0.59
# ── Server ────────────────────────────────────────────────────────────────────
fastapi==0.115.0
diff --git a/workers/rembg/requirements.txt b/workers/rembg/requirements.txt
index f7eb971b0..deaca3a62 100644
--- a/workers/rembg/requirements.txt
+++ b/workers/rembg/requirements.txt
@@ -4,7 +4,10 @@ uvicorn[standard]==0.32.1
# onnxruntime-gpu plus its CUDA libraries on top of the CPU runtime that rembg
# already installs. onnxruntime is pinned because rembg leaves it open, which
# made every rebuild a different inference runtime.
-rembg==2.0.59
+# 2.0.75 is the first release with the fixes for the model_path traversal and
+# the server SSRF/CORS advisories. It needs Python 3.11 and numpy 2.3, which this
+# python:3.11-slim image takes cleanly (onnxruntime 1.28 ships numpy 2 wheels).
+rembg==2.0.75
onnxruntime==1.28.0
# Pillow 12.3.0, not 11.x: every 11.x release carries open advisories, and this
# service decodes an image the caller supplied, which is the exact path the heap
diff --git a/workers/remesh/requirements.txt b/workers/remesh/requirements.txt
index 84ac02d5a..31a741c95 100644
--- a/workers/remesh/requirements.txt
+++ b/workers/remesh/requirements.txt
@@ -13,8 +13,10 @@ Pillow==12.3.0
google-cloud-storage==2.19.0
httpx==0.27.2
pydantic==2.10.3
-# FBX/DAE input loading via the assimp python bindings (trimesh backend).
-pyassimp==5.2.5
+# No pyassimp. trimesh 4.x carries no assimp loader at all (DAE input goes
+# through pycollada from trimesh[easy]; FBX is not a trimesh input format), and
+# nothing in this worker imports pyassimp, so the pin only shipped the 33 OSV
+# advisories filed against the assimp C++ library without ever being called.
# USD / USDZ support
# usd-core 25.11, not 24.11: 24.11 carries a CRITICAL Sdf_PathNode use-after-free
# (GHSA-58p5-r2f6-g2cj) and a file-parsing use-after-free, both RCE-class. This
diff --git a/workers/stylize/requirements.txt b/workers/stylize/requirements.txt
index 18cc99b0c..1e7ed5ad7 100644
--- a/workers/stylize/requirements.txt
+++ b/workers/stylize/requirements.txt
@@ -7,8 +7,10 @@ scipy==1.13.1
google-cloud-storage==2.19.0
httpx==0.27.2
pydantic==2.10.3
-# FBX / DAE input support via assimp python bindings
-pyassimp==5.2.5
+# No pyassimp. trimesh 4.x carries no assimp loader at all (DAE input goes
+# through pycollada from trimesh[easy]; FBX is not a trimesh input format), and
+# nothing in this worker imports pyassimp, so the pin only shipped the 33 OSV
+# advisories filed against the assimp C++ library without ever being called.
# Ed25519 signing for the OIN protocol layer (oin.py). Optional at runtime
# (pure-Python fallback), but production images take the libsodium path.
-pynacl==1.5.0
+pynacl==1.6.2
diff --git a/workers/texture/requirements.txt b/workers/texture/requirements.txt
index 4664a443f..16144b06b 100644
--- a/workers/texture/requirements.txt
+++ b/workers/texture/requirements.txt
@@ -1,8 +1,15 @@
fastapi==0.115.5
uvicorn[standard]==0.32.1
+# transformers 4.57.6 and accelerate 1.15.0: newest releases that clear their
+# advisories and still run on the torch 2.3.1 pinned below (transformers only
+# imports DTensor on torch>=2.5). diffusers stays on 0.31.0: its open advisories
+# need 0.38 and are reachable only by loading an attacker-chosen repo or
+# custom_pipeline, and this worker loads fixed SDXL/ControlNet repos from the
+# operator-staged cache. torch 2.3.1 is held by the cu121 base image and the
+# xformers 0.0.27 pairing in the Dockerfile; moving it is a base-image change.
diffusers==0.31.0
-transformers==4.47.1
-accelerate==1.2.1
+transformers==4.57.6
+accelerate==1.15.0
safetensors==0.4.5
controlnet-aux==0.0.9
trimesh[easy]==4.5.3