The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Multiple instances of remote script execution using the risky 'curl pipe bash' pattern were detected, targeting URLs such as 'https://cdn.jsdelivr.net/gh/ruvnet/ruflo@main/scripts/install.sh' and 'https://fnm.vercel.app/install'. Direct execution of unverified remote scripts bypasses local security audits.
[COMMAND_EXECUTION]: Heuristic analysis and static scanning identified code patterns consistent with fork bombs and destructive system commands (including 'rm -rf /' patterns) in core utility and hook scripts, such as 'scripts/smoke-pre-bash-hook.mjs' and '.claude/helpers/hook-handler.cjs'. The documentation also highlights historical vulnerabilities in dependencies like 'neural-trader' related to malicious install hooks.
[PROMPT_INJECTION]: The skill contains various instructions that attempt to redefine agent roles, ignore safety filters, or override previous system instructions. Significant findings include attempts to enable unrestricted modes in '.claude/commands/pair/commands.md' and override system prompts in several plugin README files.
[EXTERNAL_DOWNLOADS]: The skill facilitates the installation and execution of numerous third-party packages and remote scripts. The documentation specifically warns about the potential for malicious install scripts in its dependencies, requiring defensive measures like '--ignore-scripts' to prevent system disruption.
[DATA_EXFILTRATION]: The skill integrates analytics and telemetry tracking to an external service ('https://funnel.ruv.io/v1/events'). While described as analytics, this infrastructure creates a channel for the automated transmission of project metadata and session information to a remote server.

ruflo