Audit AI Skills for Security Vulnerabilities Cross-Platform

Audit Skills is a security auditor for AI skills and bundles. It performs non-intrusive static analysis to identify malicious patterns, data leaks, system stability risks and obfuscated payloads across Windows, macOS, Linux and Unix, and mobile (Android and iOS).

The audit runs in three steps: static analysis, platform-specific threat detection, and a report. The threat detection walks nine categories:

• Privilege, ownership and metadata manipulation: elevated access such as sudo, chown, chmod, icacls and Set-ExecutionPolicy, and metadata tampering such as touch -t, attrib and chflags. The risk is unauthorized access, masking activity or making files immutable.
• File and folder locking and resource denial: patterns such as chmod 000, chattr +i and Deny ACEs, and locking or hiding folders in user or system directories, which risks denial of service or data locking.
• Script execution and batch invocation: Windows batch and script hosts, Unix shell scripts run after chmod +x, PowerShell with execution policy bypass, and hidden window flags.
• Dangerous install, uninstall and system changes, for example silent MSI installs, package removal and deleting from system directories, which can remove security software.
• Mobile application and OS security: Android tooling such as adb shell, apktool and dex2jar, manifest and dex file manipulation, iOS tooling such as codesign and xcrun, plist and provisioning file changes, jailbreak or root detection bypasses, hardcoded API keys in mobile source, and sensitive permission requests in non-mobile skills.
• Information disclosure and network exfiltration: download and transfer tools such as curl, wget, scp and nc, access to sensitive data such as .env, .ssh, cookie stores and keychains, and scanning of internal IPs.
• Service, process and stability manipulation: stopping services, forced kills, disabling systemctl units, direct disk access with dd, firmware or BIOS calls and kernel module management.
• Obfuscation and persistence: Base64, hex and XOR encoding, atob(), Run keys, schtasks, crontab, launchctl, systemd units, and network fetches piped straight into a shell.
• Legitimacy and scope: cross-reference with CATALOG.md, check the standard repo layout, and ask whether the skill's scope justifies its commands, for example whether a UI design skill needs adb shell or sudo.

The final report carries a score from 0 to 10, the identified platform targets, flagged actions, threat analysis and mitigation recommendations.

Use it when you need to audit AI skills and bundles for security vulnerabilities, run cross-platform security analysis, verify a skill's legitimacy, or scan for mobile threats in AI skills. Do not execute potentially malicious code during the audit and do not modify the code being audited, because the method is static analysis only. Do not ignore mobile-specific concerns. It is not a substitute for environment-specific validation, testing or expert review, and it stops to ask for clarification when inputs, permissions, safety boundaries or success criteria are missing.

The input is a skill or bundle to review. The output is the scored security report described above.

"Perform a security audit on this skill bundle"

Documented pitfalls: executing code during an audit (stick to static methods), missing cross-platform threats (check every supported platform) and failing to detect obfuscated payloads (look for Base64, hex, XOR loops and atob()).

It points to @security-scanner for additional scanning capabilities, checks against the repository's CATALOG.md, and is written to work with several assistants including Claude, Gemini, GPT, Llama and Mistral.

Maintainers of skill catalogs, security reviewers and teams that install third-party AI skills and want to check them for malicious or risky behavior before use.