The Agent Skills Directory

Unverifiable Dependencies & Remote Code Execution (MEDIUM): The skill clones the BVLC/caffe repository and executes its internal scripts (get_cifar10.sh, create_cifar10.sh). Since the source is not in the trusted organizations list, this is a high-risk pattern, downgraded to MEDIUM as it is the primary purpose of the skill.\n- Privilege Escalation (MEDIUM): The skill uses apt-get to install system packages, which requires administrative access. This is downgraded from HIGH as it is essential for the framework's installation.\n- Dynamic Execution (LOW): Compiles the Caffe framework from source using make. This is a standard part of the build process but represents a surface for executing code compiled at runtime.\n- Indirect Prompt Injection (LOW): The skill processes external data from a git repository and has significant system capabilities. Evidence Chain: 1. Ingestion: git clone and data download scripts in Phase 2 and Phase 5. 2. Boundary markers: Absent. 3. Capability inventory: apt-get, make, and binary execution in Phase 7. 4. Sanitization: Absent.

caffe-cifar-10