windows-av-evasion

Fail

Audited by Gen Agent Trust Hub on Jul 21, 2026

Risk Level: HIGHPROMPT_INJECTIONREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONCREDENTIALS_UNSAFE
Full Analysis
  • [PROMPT_INJECTION]: The skill uses "AI LOAD INSTRUCTION" to direct the agent to adopt an "Expert Attack Playbook" persona focused on bypassing security controls, which aims to override standard safety constraints regarding harmful activities.
  • [REMOTE_CODE_EXECUTION]: The skill provides extensive documentation and functional code snippets for sophisticated process injection techniques, including:
  • CreateRemoteThread and NtMapViewOfSection for cross-process code execution.
  • Process Hollowing and Module Stomping to hide malicious code inside legitimate process memory.
  • Early Bird APC Injection to execute shellcode before a process's main entry point.
  • Dynamic .NET Assembly.Load patterns to run binaries entirely in memory.
  • [COMMAND_EXECUTION]: Includes specific one-liners and scripts for:
  • PowerShell version downgrading (powershell -Version 2) to bypass modern security features like AMSI.
  • Registry modifications for COM Hijacking (HKCU\Software\Classes\CLSID...) to disable security providers.
  • Memory patching of amsi.dll and ntdll.dll to suppress security alerts.
  • [DATA_EXFILTRATION]: Provides templates for fetching remote payloads from attacker-controlled infrastructure (http://attacker/payload.ps1) and methods to disable ETW (Event Tracing for Windows), which prevents security logs from recording network or file activity.
  • [CREDENTIALS_UNSAFE]: Explicitly references and facilitates the use of credential-theft tools like Rubeus and Mimikatz by providing the exact bypass chains needed to execute them on protected endpoints.
  • [INDIRECT_PROMPT_INJECTION]: The skill lacks sanitization for external content, creating a high-risk attack surface where an agent could be induced to execute malicious shellcode if it processes data (like a URL or command argument) provided by an untrusted source.
  • Ingestion points: Remote payloads via DownloadString or user-supplied execution args in SKILL.md.
  • Boundary markers: Absent.
  • Capability inventory: VirtualAlloc, CreateRemoteThread, Assembly.Load, and direct syscalls (SysWhispers).
  • Sanitization: None provided.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Jul 21, 2026, 01:24 PM
Security Audit — agent-trust-hub — windows-av-evasion