What problem does it solve? Agentic systems that chain retrieval, memory, planners, and tool calls create trust boundaries where untrusted text can hijack planner behavior, corrupt tool arguments, or expose secrets. This Skill provides a structured methodology to map those layers, prove a minimal exploit chain, and report exactly which boundary failed in CTF sandbox challenges. ## Core Features & Use Cases - Control Stack Mapping: Separates system, developer, user, retrieved, memory, planner, and tool-response layers to distinguish claimed capability from runtime-exposed capability. - Minimal Proof Chains: Reproduces one compact chain from untrusted content to changed planner behavior, altered tool arguments, or secret exposure before exploring variants. - Boundary-Based Reporting: Attributes failures to retrieval, summarizer, planner, executor, tool normalization, or output post-processing, separating instruction drift from actual side effects. - Use Case: During a CTF challenge where a retrieval-augmented agent leaks a flag, use this Skill to trace the poisoned document chunk through the summarizer into the planner and capture the final tool invocation as evidence. ## Quick Start Ask the agent to analyze the prompt injection challenge by mapping the control stack and proving one minimal chain from untrusted content to the final tool call.
