Two more OpenEnv environments are off the hook for CVE-2026-85999 and CVE-2026-86000. The soupsieve 2.8.4 to 2.9.2 bump that already landed in openapp_env now covers finrl_env and browsergym_env, with anyio moved to 4.14.2 for GHSA-82r6-8w77-94w6 and uv lock --check passing in both. TRL rerouted GRPO log-probabilities and entropy through _ChunkedLogProbFunction, feeding the trainer's native loss unchanged. That turns on off-policy and entropy options and deletes the duplicate GRPO loss files. MoE auxiliary loss still goes through the regular model forward. The PR labels it part of a larger effort,. Aria's routed experts predated the standard use_experts_implementation interface and synced to host on every call via tokens_per_expert.cpu(). They're now AriaTextTopKRouter plus AriaExperts, the same shape GraniteMoe and Qwen3-MoE use. If you touched AriaGroupedExpertsGemm or its fc1/fc2 grouping, move to the shared experts interface. In diffusers, _offload_to_disk freed onloaded tensors while the compute stream could still be reading them, so the next onload could reuse that memory and produce NaN output. It now synchronizes the compute stream the way _offload_to_memory already did. #python https://repojournal.com/showcase/huggingface/2026-09-19/openenv-closes-soupsieve-cves-in-two-more-envs-trl-reshapes-grpo-log-probs
Post summary
The OpenEnv environments finrl_env and browsergym_env have been updated via soupsieve and anyio version bumps to address CVE-2026-85999 and CVE-2026-86000, effectively patching the vulnerabilities.
