The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted binary and script samples which creates a surface for indirect prompt injection if the samples contain adversarial content.
Ingestion points: Suspicious PE, ELF, Mach-O, and script files (SKILL.md).
Boundary markers: The instructions lack explicit warnings or delimiters for the agent to treat strings and metadata within the analyzed samples as potentially malicious instructions.
Capability inventory: The agent is instructed to use a wide array of powerful tools for debugging, file system access, and network communication.
Sanitization: No specific mechanisms for sanitizing or escaping extracted sample data are defined before the agent evaluates the content.
[COMMAND_EXECUTION]: The skill instructs the agent to execute multiple command-line utilities for file triage and rule conversion.
Evidence: Use of commands such as sha256sum, rabin2, sigmac, and sigma convert throughout the methodology.
[DATA_EXFILTRATION]: The skill includes a command to transmit analysis samples to a local network service.
Evidence: The skill uses curl to upload files to a local CAPE sandbox instance at http://localhost:8000 (SKILL.md). While targeting localhost is a standard practice for isolated analysis environments, it involves the network transfer of potentially sensitive artifacts.

reverse-eng-malware