The Agent Skills Directory

EXTERNAL_DOWNLOADS (MEDIUM): The skill references a third-party GitHub repository (github.com/Svenskithesource/PyArmor-Unpacker) for unpacking protected Python binaries. This source does not belong to the trusted organizations list, making it an unverifiable dependency.
REMOTE_CODE_EXECUTION (MEDIUM): Under JavaScript Deobfuscation, the skill explicitly instructs the user/agent to 'run the obfuscated code' after replacing eval. If the agent executes this code in its own environment, it could lead to compromise, especially given that the data being processed is expected to be malicious.
COMMAND_EXECUTION (LOW): The skill provides numerous commands for interacting with potentially malicious binaries, archives, and network captures. While standard for malware analysis, these operations involve inherent risks when handling adversarial files.
INDIRECT_PROMPT_INJECTION (LOW): The skill is designed to ingest and process untrusted data (malware, PCAPs, Debian packages) using high-privilege tools like Bash and WebFetch without explicit sanitization or boundary markers between instructions and data.
Ingestion points: malware.exe, file.pcap, package.deb, and obfuscated script contents.
Boundary markers: Absent; instructions do not suggest using delimiters to separate malicious content from the prompt context.
Capability inventory: Includes Bash (shell execution), WebFetch (network access), and Write (filesystem access).
Sanitization: Absent; the skill suggests direct execution/processing of extracted payloads.

ctf-malware