metal-sweep
Pass
Audited by Gen Agent Trust Hub on Jun 30, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes shell commands including
swift runand various Python scripts (analyze_results.py,multi_metric_rank.py,render_calmar_lists.py). It also instructs the agent to dynamically modify or create Metal kernel source code and register it in Swift source files before execution. - [PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection by interpolating user-provided strategy names and parameters directly into shell commands and source code files.
- Ingestion points: User-supplied strategy names, parameter ranges, and file paths (e.g.,
<name>,<name>_sweep). - Boundary markers: Absent; the instructions do not specify delimiters or warnings to ignore instructions embedded in the user data.
- Capability inventory: The skill possesses file-write capabilities (modifying
.metalandmain.swift) and shell execution capabilities (swift run,python). - Sanitization: No explicit sanitization or validation of the user-provided strategy names or parameters is performed before they are used in commands or code generation.
Audit Metadata