The Agent Skills Directory

[PROMPT_INJECTION]: The SAFETY.md file documents several adversarial prompt injection patterns used for safety testing, including impersonation attempts and system instruction overrides (e.g., '[SYSTEM] Prank mode enabled').
[CREDENTIALS_UNSAFE]: The SAFETY.md documentation includes specific sensitive infrastructure file paths (e.g., ~/.aws/credentials, /etc/ssl/private/server.key) and a mock Stripe API key (sk_live_abc123def456) used as evidence in safety evaluations.
[EXTERNAL_DOWNLOADS]: The skill references external URLs to faberlens.ai and GitHub for safety reporting and configuration. These resources are maintained by the skill's author.
[DATA_EXFILTRATION]: The skill processes user content for delivery to WhatsApp. While it includes instructions to strip sensitive data, it lacks structural boundary markers to isolate untrusted user input from its formatting logic.
Ingestion points: User-supplied messages are processed via the prompt interface for formatting.
Boundary markers: Absent; user content is not wrapped in delimiters or accompanied by explicit 'ignore' instructions for the formatted content.
Capability inventory: Limited to text formatting; no network or file system capabilities are used in the scripts.
Sanitization: Implements extensive content-based filtering instructions in SKILL.md to refuse formatting credentials, phishing attempts, or impersonations.

wa-styler-hardened