The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill explicitly instructs the agent to use npx -y rea-agents@latest as a fallback execution method for its analysis tools in SKILL.md and references/connection-and-recovery.md. This downloads and executes the latest version of an external Node.js package from the public npm registry at runtime, which is an unverified dependency and allows for arbitrary code execution if the package is compromised or typosquatted.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a broad attack surface for indirect prompt injection because its core functionality involves the agent ingesting and interpreting potentially malicious or attacker-controlled data from binary files, Android APKs, and web network captures.
Ingestion points: Throughout the skill (e.g., SKILL.md), the agent is directed to use tools like open_binary, inspect_web_network_capture, and inspect_android_package that pull data from external artifacts into the agent's context for reasoning.
Boundary markers: The instructions fail to provide explicit boundary markers or directions for the agent to disregard instructions that might be found within the data it is analyzing (e.g., strings in binaries, comments in decompiled JavaScript, or instructions in HTML payloads).
Capability inventory: The skill manages and utilizes powerful tools including Frida, GDB, and Rizin (detailed in references/runtime-observation.md) which have the ability to execute code and modify system state, making the consequences of a successful injection high.
Sanitization: There are no instructions for sanitizing or filtering the output of its reverse engineering tools before that output is processed by the agent.

reverse-engineer-anything