The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation and examples include references to known malicious domains such as 'malicious-url.com' and 'malicious-site.com'. Automated scanners have confirmed that these domains are blacklisted, posing a risk if the agent or user accesses them directly.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from malware behavioral reports which can influence agent behavior.\n
Ingestion points: The scripts/agent.py script and SKILL.md workflows load and parse report.json files generated from malware execution stored in /opt/cuckoo/storage/analyses/.\n
Boundary markers: No boundary markers or 'ignore embedded instructions' delimiters are used when the agent processes or displays strings from these reports.\n
Capability inventory: The skill possesses network capabilities via requests to interact with the Cuckoo API and can execute shell commands like cuckoo, vol3, and curl.\n
Sanitization: There is no sanitization or validation of data extracted from the reports (e.g., process names, API arguments, or network URIs) before it is printed to the console or used in logic, creating a surface for indirect prompt injection attacks.\n- [COMMAND_EXECUTION]: The workflow documentation provides multiple shell command examples for submitting samples and performing forensics. These commands interact with local services and could lead to unintended consequences if isolation between the host and the sandbox is not strictly enforced.\n- [METADATA_POISONING]: An automated scanner has flagged references/api-reference.md as having a malicious file reputation ('FileRepMalware'), indicating that the documentation itself contains signatures associated with known threats.

analyzing-malware-behavior-with-cuckoo-sandbox