The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources and source code, creating a surface for potential indirect instructions. * Ingestion points: references/sast.md (source code, lockfiles) and references/cve.md (CVE/NVD feeds). * Boundary markers: Not explicitly defined in the provided files. * Capability inventory: Access to bash, run_command, and write_to_file. * Sanitization: Includes instructions for Vibeguard Protocol and ClawSec runtime interception to monitor output and block destructive commands.
[DYNAMIC_EXECUTION]: The skill generates and executes idempotent remediation playbooks (e.g., Ansible) based on identified vulnerabilities. This behavior is a core part of its remediation mode.
[COMMAND_EXECUTION]: The skill utilizes system-level tools to perform diagnostics and security audits. It features a dedicated runtime mode designed to intercept and block high-risk commands like rm -rf / and terraform destroy.

muse-security