ai-agent-redteam

Warn

Audited by Gen Agent Trust Hub on Jul 2, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: The orchestration script agent_redteam_harness.py executes internal testing modules via subprocess.run, allowing for automated runs of various attack scripts based on configuration files.
  • [EXTERNAL_DOWNLOADS]: Testing scripts utilize the requests library to communicate with user-provided AI agent endpoints and external OAST (Out-of-Band Application Security Testing) servers to verify vulnerabilities.
  • [PROMPT_INJECTION]: The repository includes extensive examples and automated generation scripts for prompt injection payloads, including multi-turn jailbreak techniques (Crescendo, PAIR) and indirect injection patterns.
  • [DATA_EXFILTRATION]: The skill includes tools such as indirect_injection_forge.py specifically designed to generate payloads that test for the exfiltration of sensitive data via markdown image rendering and other outbound channels.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jul 2, 2026, 01:40 PM
Security Audit — agent-trust-hub — ai-agent-redteam