The Agent Skills Directory

[Safety Refusal Taxonomy]: The skill implements a defensive safety layer by defining regex patterns for various refusal categories (jailbreak, prompt-injection, harmful-elicitation, etc.). These patterns are used to scan candidate test stimuli and refuse any that match prohibited behavior, ensuring the agent remains within policy guidelines.
[Subprocess Execution]: The import_corpus.py script and associated test files use subprocess.run to execute local PowerShell and Bash scripts. This is a standard mechanism used here to provide cross-platform support for the safety linting process and does not involve remote code execution.
[Input Sanitization]: The skill includes logic to normalize and escape user-provided inputs when generating test files. Specifically, it uses JSON-based double-quoting for YAML scalars and line-break stripping for comments, which are effective mitigations against indirect prompt injection and YAML document corruption.
[Security Testing Infrastructure]: The presence of a fuzz harness (fuzz_harness.py) and property-based tests indicates a high level of development maturity, focusing on ensuring the robustness and integrity of the data processing pipeline.

vally-tests