Fix 2 issues flagged across 2 files

[begininvoke]

A scan flagged a few things in this repository. This changes 2 files — one item each, described below.

1. eval/requirements.txt, around line 22

CRITICAL severity: Remote Code Execution in PyTorch (CVE-2025-32434). The project pins torch==2.2.0 in eval/requirements.txt (line 22), which is affected by a deserialization flaw where loading a crafted model checkpoint via torch.load() — even with weights_only=True, the historically 'safe' loading mode — allows arbitrary code execution on the host. Impact: any workflow that loads model weights from untrusted or semi-trusted sources (user uploads, public model hubs, shared artifacts, poisoned CI caches) becomes an RCE vector, potentially leading to full server compromise, credential/secret theft, and tampering with models or training data. Risk is rated CRITICAL because exploitation requires only delivering a malicious checkpoint file and it defeats the standard weights_only=True mitigation, so no code changes in the consuming application are needed for an attacker. Remediation: upgrade torch to >= 2.6.0 (patched version), regenerate any lock files, re-run the evaluation pipeline to confirm compatibility (including CUDA/torchvision matching versions), and review torch.load call sites to ensure only trusted artifacts are loaded going forward.

Upgrade torch and pillow to fixed versions to resolve critical CVEs.

For reference: rule CVE-2025-32434. Rated critical.

2. requirements.txt, around line 15

CRITICAL — Remote Code Execution in PyTorch model loading (CVE-2025-32434)

The pinned torch==2.5.1+cu121 in requirements.txt is vulnerable to a remote code execution flaw during deserialization. The key danger of this CVE is that it bypasses torch.load(..., weights_only=True), which developers have historically treated as the safe way to load model checkpoints. This means the usual mitigation pattern offers no protection.

Impact: Any code path that loads a model file from a source an attacker can influence (user-uploaded checkpoints, model registries, artifacts fetched from external storage, CI pipelines pulling third-party models) becomes an RCE vector — an attacker crafts a malicious model file, and loading it executes arbitrary code on the host with the service's privileges.

Risk level: CRITICAL (CVSS ~9.3). Exploitation is trivially scriptable, the vulnerable pattern (torch.load) is ubiquitous in ML services, and the common defense (weights_only=True) is ineffective in the affected versions.

Remediation: Upgrade to PyTorch >= 2.6.0, where the deserialization path is patched. Note: 2.6.0 does not publish +cu121 wheels — available CUDA builds are cu118/cu124/cu126 — so the CUDA variant pin (and any wheel index URL) must be updated as well. If torchvision/torchaudio are present, bump them to 0.21.0/2.6.0 for ABI compatibility.

Updates torch and pillow to versions fixing known CVEs.

For reference: rule CVE-2025-32434. Rated critical.

Take or leave whichever parts are useful. If this is not the right approach, closing is fine.

---

Found with automated scanning (RedGem) and reviewed before opening. If it is not useful, closing it is completely fine.