flydsl-kernel-authoring
Pass
Audited by Gen Agent Trust Hub on Nov 3, 2026
Risk Level: SAFECOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands for development tasks, including running benchmarks via scripts (
bash scripts/run_benchmark.sh), executing tests (python3 -m pytest), and launching GPU kernels locally. - [REMOTE_CODE_EXECUTION]: The skill provides templates for executing commands on remote hosts or inside containers using
sshanddocker exec. These patterns rely on user-supplied variables for the target host and container name, facilitating remote GPU kernel execution. - [INDIRECT_PROMPT_INJECTION]: As the skill is designed to perform code reviews and authoring tasks, it handles user-provided FlyDSL code fragments. This creates a surface for indirect prompt injection if malicious instructions are embedded within the processed code, though no active exploitation is present.
- Ingestion points: User-provided FlyDSL code fragments for review and analysis tasks (e.g., SKILL.md and evals.json).
- Boundary markers: None explicitly defined to separate user code from agent instructions.
- Capability inventory: Shell access (Bash), file editing (Edit), and file reading (Read, Grep, Glob).
- Sanitization: No specific sanitization or filtering logic is described for external code fragments.
- [DYNAMIC_EXECUTION]: The skill documents and utilizes a Just-In-Time (JIT) compilation pipeline that dynamically generates and executes GPU machine code (HSACO binaries) from Python-based Domain Specific Language (DSL) definitions.
Audit Metadata