The Agent Skills Directory

Prompt Injection (SAFE): The skill contains no malicious instructions. On the contrary, it provides detailed references for 'Jailbreak Prevention', 'Safety Patterns', and 'Hallucination Reduction' (e.g., in references/advanced-techniques.md and references/llm-prompting.md).
Data Exposure & Exfiltration (SAFE): No credentials, sensitive file paths, or unauthorized network operations were detected. Mentions of tools like google_search or code() are purely descriptive templates for model behavior.
Indirect Prompt Injection (SAFE): While the skill defines templates that ingest external data (e.g., <text>[content]</text> in references/domain-data.md), it explicitly uses and recommends XML boundary markers and 'output filtering' logic to mitigate risks.
Ingestion points: Multiple placeholders for external content across domain-specific reference files.
Boundary markers: Consistently uses XML tags (e.g., <context>, <task>, <retrieved_context>) to delimit untrusted data.
Capability inventory: None. The skill consists entirely of Markdown files and contains no executable code or scripts.
Sanitization: Recommends architectural patterns like 'Self-Critique' and 'Constitutional AI' to validate outputs.
Remote Code Execution (SAFE): No remote code patterns detected. References to frameworks like DSPy are conceptual and do not involve package installations.
Persistence & Privilege Escalation (SAFE): No system-level commands or persistence mechanisms are present.

ai-artist