kernel-triton-writing
Pass
Audited by Gen Agent Trust Hub on Oct 22, 2026
Risk Level: SAFEDYNAMIC_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The helper scripts scripts/verify_kernel.py and scripts/benchmark_kernel.py use the importlib library to dynamically load and execute the Python module generated by the agent. This allows the agent to run and test the code it creates.
- [COMMAND_EXECUTION]: Both utility scripts utilize subprocess.run to execute temporary Python harness scripts. These harnesses are responsible for running the kernel and gathering performance or correctness data.
- [INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it generates executable code based on user-provided operator descriptions. This generated code is then executed locally during the verification phase. A malicious user could craft a request that tricks the agent into generating a kernel with harmful side effects. 1. Ingestion points: User requests for kernel implementation are processed during the design and writing phases of the workflow. 2. Boundary markers: No specific boundary markers or instructions to ignore embedded commands are present in the documentation. 3. Capability inventory: The scripts in the scripts/ directory provide the capability to load Python modules and execute shell commands. 4. Sanitization: There is no mechanism to sanitize or verify the safety of the agent-generated code before it is executed.
Audit Metadata