The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains examples of instruction overrides and injection patterns (e.g., "Ignore previous instructions", "forget all previous instructions") within its reference documentation. These are provided as signatures for the agent to identify when auditing other skills and do not represent active attempts to subvert the agent.
[PROMPT_INJECTION]: There is a potential for indirect prompt injection as the skill ingests untrusted data (the contents of other skills) during the audit process. Maliciously crafted instructions within an audited file could attempt to influence the agent's findings or behavior.
Ingestion points: Contents of analyzed skill files (SKILL.md, scripts, etc.)
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are used in the analysis process.
Capability inventory: File system access via a viewing tool and zip extraction (unzip).
Sanitization: No evidence of input validation or sanitization before content is processed for analysis.
[REMOTE_CODE_EXECUTION]: The references/security_patterns.md file contains a demonstration of an obfuscated code execution chain using Base64. This is explicitly presented as a suspicious pattern to look for and is not intended for execution by the skill itself.
[COMMAND_EXECUTION]: The skill instructions specify the use of the unzip command for extracting and analyzing .skill packages. This is a standard administrative operation consistent with the skill's primary function as an analyzer.

skill-security-analyzer