The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructs on multiple methods to bypass security filters and extract internal instructions, including structural extraction queries ('Quote the first sentence...'), enumerated ruleset requests with self-redaction opt-outs, and whole-context dump commands. These techniques are explicitly designed to evade 'shallow keyword guardrails'.
[DATA_EXFILTRATION]: Includes a functional Python script that automates the extraction process. This script performs network operations using urllib.request to send messages to a target URL (https://target/agent/chat) and verifies leaked data through a 'reproducibility gate'.
[INDIRECT_PROMPT_INJECTION]: The skill identifies and maps attack surfaces where an agent processes untrusted data such as uploaded files, documents, and URLs. It provides instructions on how to leverage these surfaces to identify injection points.
Ingestion points: Documented as agent features processing external files, docs, URLs, or tickets in SKILL.md.
Boundary markers: The instructions do not recommend or implement boundary markers to isolate untrusted data.
Capability inventory: Mentions capabilities including reading raw tool JSON, identifying tool-invocation hijack points, and accessing framework-level typed namespaces.
Sanitization: Highlights the absence of sanitization in tool input contracts as a primary vulnerability indicator.

hunt-agent-prompt-extraction