The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes every file within a project, including source code, configuration, and documentation, which constitutes a surface for indirect prompt injection from malicious content inside the project being audited.
Ingestion points: Step 1 and Step 3 involve a full inventory and manual review of all project files without extension filters.
Boundary markers: The skill lacks explicit data/instruction delimiters but uses meta-instructions like "Never fake a scan" and "Fix the actual issue" to ensure the agent maintains objective focus.
Capability inventory: The skill possesses the capability to execute command-line tools and modify project files to apply fixes.
Sanitization: No specific sanitization for the content of the files being read is defined, though the instructions for the agent emphasize verification through official tools.
[COMMAND_EXECUTION]: The skill instructs the agent to execute various security-related CLI tools as part of its primary function.
Findings: Mentions executing npm audit, pip-audit, cargo audit, bandit, and semgrep.
Context: These executions are standard for security auditing and the skill explicitly mandates that tools must be sourced from official registries or repositories.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill accesses sensitive local files to identify and remediate hardcoded credentials and configuration weaknesses.
Findings: Targets .env files, infrastructure-as-code files, and CI/CD configurations.
Context: Access is restricted to the local filesystem for the purpose of removing secrets; no external exfiltration patterns or unauthorized data transfers were detected.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill allows for the installation of security tools from external sources.
Findings: References tools like gitleaks, trufflehog, and osv-scanner.
Context: The skill includes strong "Trust boundaries" that prohibit installing obscure tools or piping scripts from unverified URLs, strictly requiring official package registries or GitHub releases.

security-hardening-wizard