The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, such as malware sandbox reports, community-sourced IOCs, and hash lookup results, which constitutes a potential surface for indirect prompt injection. • Ingestion points: Data enters the agent's context through external sandbox reports (CAPE, ANY.RUN, Joe Sandbox), VirusTotal lookups, and the analysis of suspicious files mentioned in the intake phase. • Boundary markers: The instructions do not define specific delimiters or instructions for the agent to ignore potentially adversarial content embedded within the analyzed report text. • Capability inventory: The skill description focuses on analytical workflows and does not include scripts that perform file system modifications, subprocess execution, or network write operations. • Sanitization: There are no specific instructions for escaping or validating the content of external reports before the agent processes them.
[SAFE]: The skill includes strong security guidance, specifically emphasizing a 'Sandbox-only discipline' and warning against the detonation of samples in production environments or on workstations with network access.
[SAFE]: All external references provided in the skill target well-known and reputable security organizations and services, such as MITRE ATT&CK, VirusTotal, Hybrid-Analysis, and established open-source repositories for malware signatures and analysis tools.

malware-triage