Building an Indirect Prompt Injection Workflow

This technical blog documents an automated methodology for crafting and testing indirect prompt injection payloads against agentic LLM systems (Claude Sonnet 4.5/4.6) using an automated Codex-driven workflow. The author demonstrates stages from simple echo tests to full chain attacks including base64 encoding, remote HTTP exfiltration, and execution of system commands (whoami), discusses model-specific defenses and evasions, and packages the approach as a reusable 'enum-indirect-prompt-injection' skill for iterative payload generation and analysis.