The Agent Skills Directory

[PROMPT_INJECTION]: Static analysis flagged several reference files (e.g., references/19-security.md, references/20-purview.md, and references/22-design-patterns-core.md) for containing prompt injection patterns such as 'ignore previous instructions' and 'DAN'. However, these strings are documented exclusively as part of security configuration examples and 'Jailbreak Detection' patterns intended to be detected and blocked by users' implementations. They are educational/defensive content and do not represent actual injection attempts.
[EXTERNAL_DOWNLOADS]: The skill includes a utility script scripts/check-updates.py that performs a GET request to the PyPI registry (https://pypi.org) to verify the latest version of the framework. Additionally, documentation in references/10-providers.md provides standard installation commands for Ollama (https://ollama.ai). These are well-known and trusted technology services.
[INDIRECT_PROMPT_INJECTION]: The framework documentation identifies potential vulnerability surfaces where untrusted data enters the agent context (e.g., in references/11e-workflow-agents.md). It explicitly provides mitigation guidance, such as boundary markers and sanitization practices, to defend against indirect injection in production systems.
[REMOTE_CODE_EXECUTION]: While setup instructions in references/10-providers.md mention piping a remote script to a shell for Ollama installation, this is provided as documentation for a well-known third-party service and is not executed automatically by the skill or the agent logic.

ms-agent-framework