huggingface-spaces

HIGH W007: Insecure credential handling detected in skill instructions.

• Insecure credential handling detected (high risk: 0.80). The skill instructs users/agents to pass tokens or secrets directly on the command line (e.g., --token, --secrets KEY=val / hf spaces secrets set ...), which would require inserting secret values verbatim into commands and could cause secrets to be exposed or exfiltrated.

CRITICAL E005: Suspicious download URL detected in skill instructions.

• Suspicious download URL detected (high risk: 0.90). These URLs include direct .whl binaries served from individual GitHub release pages (unknown/user-hosted accounts like LDYang694 and adithyaxx) which are high-risk channels for distributing executable packages compared to official vendor/CDN or vetted Hugging Face wheels.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The skill includes explicit runtime install/fetch patterns that pull and execute remote code/wheels (e.g. pip/git installs and direct .whl URLs such as git+https://github.com/graphdeco-inria/diff-gaussian-rasterization.git and https://github.com/adithyaxx/flash-attention/releases/download/v2.8.3/flash_attn-2.8.3+cu13torch2.11cxx11abiTRUE-cp312-cp312-linux_x86_64.whl), which are executed at Space startup / module import and are required dependencies for those deployment patterns.