Security

This repository covers prompt and agent risks without adding harmful jailbreak libraries or leaked proprietary prompt collections.

Covered areas:

• Direct prompt injection.
• Indirect prompt injection.
• Tool abuse.
• Data exfiltration.
• Secret exposure.
• Excessive agency.
• Untrusted retrieved content.
• Output validation.
• Human approval.
• Least privilege.
• Sandboxing.
• Audit trails.

The baseline control is simple: separate untrusted content from instructions, limit tool authority, validate outputs before use, and require human approval for sensitive actions.