The Agent Skills Directory

[COMMAND_EXECUTION]: The skill contains a vast array of high-risk commands for shell and PowerShell environments. This includes the use of exploitation tools like mimikatz, secretsdump.py, psexec.py, and responder to perform network-based attacks and local credential extraction.
[CREDENTIALS_UNSAFE]: The instructions detail how to extract sensitive credentials, including NTLM hashes, Kerberos TGT/TGS tickets, and cleartext passwords from various sources such as Group Policy Preferences (GPP), LAPS, and GMSA. Tools like Get-GPPPassword.py and secretsdump.py are specifically used for this purpose.
[PERSISTENCE]: The skill provides explicit instructions for maintaining long-term access to a compromised domain. This includes forging Golden Tickets (using mimikatz or ticketer.py) and using SharpGPOAbuse.exe to create backdoor user accounts or scheduled tasks through Group Policy Objects.
[PRIVILEGE_ESCALATION]: The primary focus of the skill is to escalate privileges from a standard domain user to a Domain Administrator. It leverages techniques such as Kerberoasting, AS-REP Roasting, DCSync, and exploit scripts for critical vulnerabilities like ZeroLogon (CVE-2020-1472) and PrintNightmare (CVE-2021-1675).
[INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection as it ingests and processes untrusted data from Active Directory objects and tool outputs without sanitization.
Ingestion points: Tool outputs from BloodHound, PowerView, and LDAP queries (e.g., Get-NetUser, Get-DomainObjectAcl) containing attributes controlled by potentially malicious actors in the environment.
Boundary markers: Entirely absent; the agent is instructed to process and act on these outputs directly.
Capability inventory: Subprocess execution (bash, powershell, python3), file system writes, and extensive network operations.
Sanitization: No evidence of escaping or validation of output before it is used to form subsequent commands or agent decisions.

active-directory-attacks