The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions and commands to execute untrusted malware samples locally (e.g., using strace on the $SAMPLE_NAME). While the skill includes warnings about using isolated environments, an automated agent executing these commands on a production host would lead to system compromise.
[DATA_EXFILTRATION]: The workflow involves sending untrusted files and their hashes to external services, including VirusTotal and a user-configured sandbox URL. This is standard for malware analysis but involves external network transmission of local data.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection. It extracts and displays strings, metadata, and reports from untrusted malicious samples. If a sample contains crafted text intended to manipulate the agent (e.g., within extracted strings or macro analysis), the agent may process these as valid instructions.
[EXTERNAL_DOWNLOADS]: The skill encourages the installation of several external tools and Python packages (e.g., yara-python, pefile, oletools, floss, and die). While these are well-known security tools, they involve downloading and executing code from external repositories.
[CREDENTIALS_UNSAFE]: The skill attempts to source credentials from ~/.config/security-skill/secrets.env and handles a VT_API_KEY provided via environment variables. While this follows common practices for tool configuration, it involves managing sensitive API keys.

malware-analysis