The Agent Skills Directory

[SAFE]: The skill utilizes local TypeScript scripts (scripts/run-eval.ts and scripts/review-eval.ts) to orchestrate the evaluation workflow. These scripts include robust path validation mechanisms, such as safeChild and assertCanonicalWithin, which effectively prevent directory traversal attacks by ensuring file operations remain within expected directories.
[SAFE]: When generating text candidates via external CLIs (Codex or Claude Code), the skill explicitly disables high-risk capabilities like shell access, web search, and image generation. This follows the principle of least privilege, significantly reducing the attack surface for prompt injection or data exfiltration.
[SAFE]: A specialized validation function, assertNoLocalImagePath, is implemented to scan all input data for local file paths. This prevents potential exploits targeting model-specific tools (like view_image) that might otherwise be used to access sensitive local files.
[SAFE]: Untrusted data from evaluation cases is safely interpolated into prompts using JSON.stringify and is wrapped in explicit instructions that command the model to treat the content strictly as data, mitigating the risk of indirect prompt injection.

evaluate-ghostwriter