The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes a catalog of adversarial prompt patterns, such as instructions to ignore previous rules or reveal system prompts, within a RED_TEAM_PROMPTS list and a Python-based detection script. These are provided strictly for testing and monitoring purposes to improve agent robustness.- [REMOTE_CODE_EXECUTION]: Guidance is provided for installing the gVisor sandbox via curl from its official domain. This is presented as a best practice for isolating untrusted tool execution in a secure runtime environment.- [OBFUSCATION]: Examples of Base64 encoding and hidden Unicode characters are used to illustrate common evasion techniques. These examples are part of a security testing suite used to verify the effectiveness of input filters.- [DATA_EXFILTRATION]: The skill defines regex patterns for sensitive data types like AWS keys and connection strings. These patterns are used in a redaction utility designed to detect and block accidental data leakage in agent responses.- [COMMAND_EXECUTION]: Security configurations for Docker and gVisor are documented, demonstrating how to use least-privilege principles, such as dropping capabilities and using read-only filesystems, when executing agent tools.

ai-agent-security