The Agent Skills Directory

[SAFE]: The skill serves as an architectural guide for developing secure AI agent infrastructures. It does not contain executable code or instructions that would compromise agent safety.
[EXTERNAL_DOWNLOADS]: The documentation references official resources and research from well-known and trusted organizations, including Anthropic, OpenAI, Google, Microsoft, and Databricks. These references are used for educational and engineering purposes.
[COMMAND_EXECUTION]: While the documents discuss shell and file system access, they do so in the context of designing secure sandboxes and isolation mechanisms (e.g., using OS primitives like bubblewrap and gVisor) to prevent unauthorized execution.
[CREDENTIALS_UNSAFE]: The skill provides explicit guidance on how to avoid hardcoding credentials, recommending the use of secure vaults and proxies. It mentions sensitive paths like ~/.aws/credentials only as examples of data that should be protected by security boundaries.
[PROMPT_INJECTION]: The content analyzes prompt injection as a threat vector and provides architectural remediations, such as reasoning-blind classifiers and structural markers for untrusted content. It does not contain any malicious injection attempts.
[DATA_EXFILTRATION]: The skill identifies data exfiltration as a primary risk for agents and recommends network egress controls and domain allowlisting as defensive measures.
[INTERNAL_RESOURCES]: The references to paths under ~/ghq/github.com/efoo-team/ are identified as vendor resources associated with the author 'efoo-team', representing their internal engineering documentation and architectural decisions.

agent-harness-engineering