The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains keywords often associated with prompt injection (e.g., 'ignore previous instructions', 'admin mode', 'unrestricted mode'). However, these are documented as examples of threats to look for during a security audit or as part of a 'Prompt Injection Safety Notice' that instructs the agent on how to handle malicious content in target files. These are benign and defensive in this context.
[EXTERNAL_DOWNLOADS]: The skill references several reputable security frameworks and research papers (OWASP, NIST, MITRE). It also includes a reference to an open-source exploit playground on GitHub (fabraix/playground). These references are used for informational and assessment purposes and do not involve automated execution of untrusted code.
[DATA_EXFILTRATION]: The skill instructs the agent to search for sensitive patterns such as API keys and credentials. These operations are restricted to the local environment and are part of the intended security review process to identify hardcoded secrets in the target application's source code.
[COMMAND_EXECUTION]: The skill uses read-only tools (Read, Grep, Glob) to inspect the target system. It does not contain instructions for executing arbitrary shell commands or modifying the host system.
[SAFE]: The skill implements security best practices by including an explicit 'Prompt Injection Safety Notice' section. This section provides boundary markers and instructions to treat all analyzed content as untrusted data, mitigating risks associated with indirect prompt injection during the review process.

agentic-top-10