The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides multiple commands for downloading and immediately executing remote content from arbitrary sources. Evidence includes piping curl output to bash (curl ... | bash) and utilizing PowerShell's Invoke-Expression (IEX) to run remote scripts. Automated scanners specifically identified multiple untrusted RCE patterns involving piped execution and downloaded script execution.
[DATA_EXFILTRATION]: Instructions include techniques for exfiltrating sensitive data such as /etc/passwd through multiple protocols including HTTP POST, SMB, and FTP. It further details covert channels using DNS queries and ICMP packets to tunnel data out of a network. The skill references external resources like HackTricks for advanced exfiltration methodologies, which are flagged as malicious by reputation scanners.
[OBFUSCATION]: The skill documents techniques to evade security controls, such as Base64 and Hex encoding for binary data, and string manipulation in PowerShell (I[E]X) and Windows LOLBAS commands (RunHTML[A]pplication) to bypass signature-based antivirus detection.
[DYNAMIC_EXECUTION]: Features methods for runtime code loading, specifically reflective DLL loading using [System.Reflection.Assembly]::Load which executes payloads directly in memory, and the generation of VBS downloaders on the fly via cmd.exe echo commands.
[PRIVILEGE_ESCALATION]: Includes commands that require or grant elevated privileges, such as using sudo for network listeners and reconfiguring database engines to enable operating system command execution via xp_cmdshell in MSSQL.
[PERSISTENCE]: The skill explicitly documents "Persistence and Staging" techniques, providing the delivery mechanisms necessary to establish and maintain unauthorized access across sessions through remote code execution.
[COMMAND_EXECUTION]: Provides a wide array of commands for system interaction and tool movement across Linux and Windows, leveraging native binaries (LOLBAS/GTFOBins) to execute arbitrary code and blend in with legitimate system activity.
[EXTERNAL_DOWNLOADS]: Includes numerous methods for fetching tools and payloads from remote servers using wget, curl, certutil, and bitsadmin.
[INDIRECT_PROMPT_INJECTION]: The skill creates a significant attack surface by taking URLs and file paths from user input and processing them with powerful utilities. There are no boundary markers or sanitization steps to prevent embedded instructions in the transferred data from being executed by the agent.
Ingestion points: Processes URLs and remote file content supplied in prompts or retrieved from external sources.
Boundary markers: Absent. The instructions do not include delimiters or warnings to ignore instructions embedded in the data being transferred.
Capability inventory: Extensive capabilities including shell access, network operations, and file system writes across both Linux and Windows platforms.
Sanitization: Absent. There is no evidence of input validation or escaping for the URLs or data processed by the transfer tools.

transferring-files