This skill is a downstream specialization of the ctf-sandbox-orchestrator skill. It covers challenges about trust boundaries inside an agentic system: prompt injection, retrieval poisoning, memory contamination, planner drift, tool-argument corruption and secret exposure.
The agent maps the layers (system, developer, user, retrieved, memory, planner, tool response), then documents one minimal chain from untrusted text to a changed tool call or exposed secret, in compact evidence blocks. Replies are in Simplified Chinese unless English is requested.
What You Get
- A layer map separating system, developer, user, retrieved, memory, planner and tool-response content
- A minimal proof chain from untrusted content to an observable side effect
- A reporting rule that names the failing layer: retrieval, summarizer, planner, executor or post-processing
- Three compact evidence blocks to keep: the source chunk, planner drift and final tool call
Common Use Cases
- Explaining how a poisoned document changed an agent’s tool arguments in a lab
- Reviewing memory contamination or planner drift in a sandbox agent chain
- Writing up a secret exposure caused by an agent tool boundary in a CTF
What Makes It Unique
It treats a malicious string as insufficient proof until a side effect is observed, and it asks the reporter to name which layer failed. It is limited to agent trust boundaries and runs only after the orchestrator skill has set sandbox assumptions.
Under The Hood
The skill is instruction text only: a SKILL.md, one reference checklist and an agent config file that disables implicit invocation. It ships no scripts and does not itself fetch pages, call APIs, run commands or write files. Any observation of the challenge agent uses the tools the user already provides in the sandbox.