The Agent Skills Directory

Prompt Injection (SAFE): The skill focuses on legitimate task orchestration. No bypass markers, jailbreak attempts, or system prompt extraction patterns were found.
Data Exposure & Exfiltration (SAFE): Explicit guardrails in references/CHECKLISTS.md and references/RUBRIC.md forbid requesting or storing secrets and credentials. No patterns for exfiltrating sensitive local files (e.g., ~/.ssh, .env) were observed.
Obfuscation (SAFE): No Base64, zero-width characters, homoglyphs, or encoded commands were detected across any of the files.
Command Execution & Remote Code Execution (SAFE): While the skill involves generating and running code, it enforces a 'least privilege' model and requires confirmation for risky actions. It does not contain pre-defined malicious remote script downloads.
Indirect Prompt Injection (SAFE):
Ingestion points: User-provided "Prototype goal" and "Data/integrations" in SKILL.md.
Boundary markers: Prompt templates in references/TEMPLATES.md define structured interactions to separate intent from execution.
Capability inventory: File writing and command execution are performed during the "Scaffold" and "Iterate" steps.
Sanitization: The workflow requires manual verification and confirmation gates for risky operations.
Dynamic Execution (SAFE): The skill's primary purpose is prototyping, which involves generating code from known templates. This behavior is expected and mitigated by strong safety instructions.

vibe-coding