The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The file test-pressure-3.md contains a hardcoded GitHub token (ghp_aB12cD34eF56gH78iJ90kL12mN34oP56qR78). Although found within a pressure-test scenario file, hardcoded credentials with known prefixes are classified as high risk.
[COMMAND_EXECUTION]: SKILL.md utilizes several shell commands (including ruby, bundle, grep, ls, and cat) to detect the software environment and project configuration.
[DATA_EXFILTRATION]: The skill collects source code excerpts, stack traces, and session history into a single artifact. While intended for user-controlled export to external models, this process centralizes sensitive data for egress.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by interpolating untrusted file contents into a prompt template for another AI model.
Ingestion points: Files and debug output are gathered in SKILL.md (Steps 2 and 4).
Boundary markers: File content is enclosed in markdown code blocks in template.md.
Capability inventory: The skill performs shell command execution (Step 1) and file system writes (Step 6).
Sanitization: While secret redaction is implemented in Step 5, there is no logic to sanitize or escape project content to prevent instruction injection attacks against the recipient model.

consulting-an-oracle