cuda
Pass
Audited by Gen Agent Trust Hub on Feb 17, 2026
Risk Level: SAFE
Full Analysis
- Prompt Injection (SAFE): No evidence of instructions designed to override agent behavior or bypass safety filters was found. The documentation is strictly technical and informative.
- Data Exposure & Exfiltration (SAFE): The skill does not contain any hardcoded credentials, API keys, or logic that accesses sensitive system files. All data is public domain documentation.
- Obfuscation (SAFE): No use of Base64 encoding, zero-width characters, or hex escapes was detected within the documentation text to hide commands or instructions.
- Unverifiable Dependencies & Remote Code Execution (SAFE): The skill does not include package management files (like package.json or requirements.txt) or scripts that download and execute external code.
- Command Execution (SAFE): Documentation regarding debugging and profiling tools (Nsight Compute, Nsight Systems) contains command-line examples, but these are for informational purposes and do not involve automated or hidden execution of commands.
- Indirect Prompt Injection (SAFE): While the skill provides a large context of documentation, it does not possess any dangerous capabilities (like network access or arbitrary file writing) that would create an exploitable surface for indirect prompt injection from data it might process.
- Dynamic Execution (SAFE): No use of eval(), exec(), or runtime compilation techniques was detected. The files are static markdown references.
Audit Metadata