The Agent Skills Directory

[SAFE]: All instructional content, shell command examples (e.g., curl | bash), and sensitive file references (e.g., ~/.aws/credentials, .env) are used purely for educational purposes to describe security vulnerabilities and best practices. No malicious code or hardcoded credentials were found.
[SAFE]: The automated detection of prompt injection patterns (e.g., "Ignore previous instructions") is a false positive. These phrases are contained within the llm-and-ai-security.md reference file as part of the definition of prompt injection, not as instructions intended to override the agent's behavior.
[INDIRECT_PROMPT_INJECTION]: The skill has an ingestion surface for untrusted data when inspecting system designs and dependencies as part of its primary workflow.
Ingestion points: Data enters the agent context through the inspection of system designs, data flows, and third-party dependencies as described in SKILL.md.
Boundary markers: The skill explicitly includes warnings in SKILL.md to treat all tool results and third-party artifacts as untrusted and to enforce authorization outside the model.
Capability inventory: The primary capabilities are generating reports and findings; no autonomous execution of high-privilege commands or file-writing is initiated by the skill itself.
Sanitization: The skill references external frameworks like OWASP Top 10 for LLMs and provides guidance on output filtering and schema validation to sanitize model interactions.
[NO_CODE]: The skill consists entirely of markdown documentation and guidance; no executable scripts, binaries, or configuration files that trigger runtime actions were provided.

using-security-architect