hunt-agent-tool-injection
Pass
Audited by Gen Agent Trust Hub on Nov 2, 2026
Risk Level: SAFE
Full Analysis
- [COMMAND_EXECUTION]: The skill contains a wide variety of shell command injection payloads (e.g., using separators like ';', '&&', and '$()') and bypass techniques (e.g., newline characters and character encoding). These are documented as reference material for security professionals to use when auditing third-party agents and do not pose a risk to the host environment.
- [DATA_EXFILTRATION]: Provides instructional content on how to perform data exfiltration from vulnerable agents, including the use of Out-of-Band (OOB) channels like DNS and HTTP to bypass blind injection scenarios. It specifically highlights high-value targets like cloud metadata endpoints (169.254.169.254) and IAM credentials.
- [DYNAMIC_EXECUTION]: Includes a 'Runnable Probe' Python script that facilitates vulnerability testing by measuring response times or triggering external callbacks. The script uses standard libraries (urllib.request) and acts as a template, requiring user-provided targets and credentials to function.
- [PROMPT_INJECTION]: Discusses methods for bypassing agent-level security filters and sanitization logic, such as the 'newline bypass' and exploiting differences between browser-based and API-based tool implementations.
Audit Metadata