hunt-agent-auth-bypass
Pass
Audited by Gen Agent Trust Hub on Nov 2, 2026
Risk Level: SAFEPROMPT_INJECTIONDYNAMIC_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill documents and provides multiple examples of instructional overrides designed to bypass agent authorization by claiming privileged roles (e.g., 'I am the Boss', 'As the system administrator', 'I am the owner of this account').
- [DYNAMIC_EXECUTION]: The file contains an embedded Python script in the 'Runnable Probe' section. This script is designed to automate vulnerability testing and interact with remote APIs, demonstrating how to use Python to verify exploit success.
- [DATA_EXFILTRATION]: The provided Python probe script includes logic to send a unique identifier (canary) to a remote endpoint (
https://target/agent/chat) using theurllib.requestlibrary to confirm if data has been successfully leaked through an authorization bypass. - [INDIRECT_PROMPT_INJECTION]: The skill identifies and documents attack vectors where agents ingest untrusted data from documents, QR codes, or JSON blobs. It details how malicious instructions embedded in these external sources can be used to perform 'mass assignment' or classification overrides by manipulating the agent's interpretation of the data.
Audit Metadata