The Agent Skills Directory

[PROMPT_INJECTION]: The skill references common injection strings like 'ignore previous instructions' and 'you are now...', but these are explicitly used within defensive checklists and detection logic. These instructions teach the agent how to recognize and flag malicious prompts from external sources rather than attempting to subvert the agent's own programming.
[COMMAND_EXECUTION]: The skill includes a local utility, scripts/security-audit.sh, which is designed to scan for exposed secrets and verify file permissions within the user's workspace. These operations are local, transparent, and intended to harden the security environment of the agent.
[CREDENTIALS_UNSAFE]: Security hardening rules in the skill enforce the use of a dedicated .credentials/ directory that is strictly gitignored. The audit script verifies that these files have restricted permissions (chmod 600) to prevent unauthorized local access, aligning with industry best practices for secret management.
[INDIRECT_PROMPT_INJECTION]: The framework explicitly recognizes external data (emails, websites, PDFs) as untrusted. It implements instruction isolation by directing the agent to classify such content as data only and to never execute instructions found within external documents, effectively mitigating the risk of indirect injection attacks.
[DYNAMIC_EXECUTION]: The skill performs routine system maintenance and security audits using pre-defined local scripts and standard system binaries. There is no evidence of the skill downloading, generating, or executing unverified code from remote or untrusted sources.

proactive-agent