training-mfu-calculator
Pass
Audited by Gen Agent Trust Hub on Oct 10, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill suggests running hardware diagnostic tools such as
npu-smi infoandnvidia-smito obtain peak FLOPs information from the training environment. These are standard operations for performance analysis tasks. - [INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest architectural parameters and performance data from external sources such as
config.json, training scripts, and log files. This creates a surface for indirect prompt injection if the processed files are attacker-controlled. - Ingestion points:
config.json, shell scripts, and training logs as described inSKILL.md. - Boundary markers: The instructions do not specify the use of delimiters or ignore-instructions when reading external files.
- Capability inventory: The skill's logic is implemented in
mfu_calculator.py, which is limited to mathematical calculations and does not include network, file-write, or subprocess capabilities. - Sanitization: The logic for extracting step times from logs uses regular expressions followed by floating-point conversion, which provides inherent validation for that specific numerical data field.
Audit Metadata