The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external vulnerability data retrieved via the strale_cve_lookup tool, which creates a surface for indirect prompt injection. Malicious actors could theoretically embed instructions in CVE descriptions or security advisories to influence the agent's behavior during analysis.
Ingestion points: External security data and vulnerability descriptions are fetched through the strale_cve_lookup tool, as referenced in SKILL.md and references/sandbase-api-map.md.
Boundary markers: The instructions do not provide explicit delimiters or "ignore instructions" prompts to isolate the external data from the agent's primary instructions.
Capability inventory: The skill utilizes the sandbase_call_tool gateway for API interactions (SKILL.md). While the guidelines mandate "Read-only research only," the agent processes data that informs its subsequent synthesis and reporting steps.
Sanitization: The skill lacks specific instructions for sanitizing or filtering the content of the retrieved vulnerability data before it is incorporated into the agent's context.

cve-lookup