Security audit
Security checks for vulnerabilities and agentic risk
This evaluation skill is mostly coherent, but it can commit reports or create tracker issues without clear consent or redaction guidance, which could expose sensitive prompts or outputs.
Install only if you are comfortable with the agent producing evaluation artifacts and potentially using connected Git or project-tracking tools. Before allowing commits or issue creation, choose the destination explicitly and redact secrets, credentials, private prompts, personal data, and raw transcripts from reports and reproduction steps.
The skill explicitly instructs submitting test cases and reports to a Git repository but does not require sanitizing prompts, model outputs, logs, or user-provided data first. In an evaluation skill, those artifacts commonly contain proprietary prompts, credentials, PII, or sensitive failure traces, so this creates a realistic data disclosure risk through source control exposure.
The instruction to create bug Issues with reproduction steps can leak sensitive inputs, outputs, stack traces, or internal prompts into a project tracker without any warning or redaction step. Because repro steps are often copied verbatim from failing tests, this skill context makes accidental disclosure especially plausible.
Stating that outputs are written to a local Markdown file by default introduces data persistence without informing users that reports may contain sensitive prompts, test inputs, or model outputs. Local storage is lower risk than publishing to external systems, but it still increases exposure through shared machines, backups, or later accidental redistribution.
This markdown file contains user-facing natural language and examples exclusively in Chinese, beginning with the title and continuing throughout the template. Under the policy rule, forcing a specific language without user opt-in can be a locale-policy violation when no alternative language option or justification is provided.
