The Agent Skills Directory

[PROMPT_INJECTION]: Indirect Prompt Injection Surface.
Ingestion points: SKILL.md (lines 78-85) and references/dvm-kinds.md (lines 62-67) describe processing input from url, event, and job types.
Boundary markers: The code examples for service providers in SKILL.md do not include delimiters or specific instructions to the processing LLM to ignore embedded commands within the fetched data.
Capability inventory: The skill demonstrates network read capabilities (fetch) and network write capabilities (publishing Nostr events) following the processing of untrusted content.
Sanitization: No input validation or content filtering is shown in the processing workflow.
[DATA_EXFILTRATION]: Unvalidated Network Requests.
The skill instructs developers to use fetch() on URLs provided in the Nostr i tags. This pattern enables server-side request forgery (SSRF), where an attacker could provide a URL that forces the agent to probe internal network services or access sensitive cloud metadata endpoints.

nostr-dvms