The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. It instructs the agent to read and analyze cuTile source code (e.g., source.py) to perform conversions. Adversarial instructions hidden in the source code (e.g., within comments or metadata) could potentially trick the agent into misusing its granted tools.
Ingestion points: The agent reads cuTile source files as part of the 'Pre-flight Analysis' and 'Convert' phases.
Boundary markers: There are no explicit instructions or delimiters provided to the agent to treat the source code as untrusted data or to ignore embedded instructions.
Capability inventory: The skill allows the use of Read, Write, Grep, Glob, and Bash tools, providing a significant capability surface for an injection to exploit.
Sanitization: No sanitization or validation of the input source code is performed by the skill's instructions.
[COMMAND_EXECUTION]: The skill's workflow explicitly requires the agent to execute shell commands via the Bash tool for testing and benchmarking. Examples include python -m pytest and grep commands for code analysis. While these are necessary for the skill's primary purpose of kernel conversion and validation, they represent a capability that requires careful monitoring during execution to prevent unauthorized command injection or accidental destructive behavior.

tilegym-converting-cutile-to-triton