The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from threat actor forum posts and CTI reports as part of its core workflow. This creates a potential surface for indirect prompt injection where instructions could be embedded in the analyzed intelligence.
Ingestion points: Workflow step 1 in SKILL.md describes extracting metadata from external reports, blogs, or claims.
Boundary markers: The skill does not explicitly define delimiters for untrusted input, though it instructs the agent to extract structured fields which provides inherent compartmentalization.
Capability inventory: The skill does not define or use any dangerous tools such as shell execution, file system modification, or non-whitelisted network requests.
Sanitization: No specific sanitization or filtering of input content is mandated before the assessment phase, which is typical for specialized analysis tools.
[SAFE]: The skill uses defanged URLs (e.g., Exploit[.]in, hxxps://...) in its examples and templates. These are standard security industry practices for documenting potentially malicious infrastructure safely and do not constitute malicious obfuscation, hidden links, or command injection attempts.

admiralty-system