The Agent Skills Directory

COMMAND_EXECUTION (HIGH): The skill provides templates for executing high-privilege commands using vercel, stripe, and npx convex on production environments. Specifically, the use of npx convex run --prod and vercel logs --prod with interpolated variables like [query], [app], and [pattern] poses a severe risk of shell injection if the inputs are not strictly validated.
DATA_EXFILTRATION (MEDIUM): The protocol encourages reading production logs and querying specific database records using user-supplied criteria. In an adversarial scenario, an attacker could craft inputs to expand the scope of these queries to expose sensitive data from the production database or logs.
INDIRECT_PROMPT_INJECTION (HIGH): (Category 8) Evidence Chain: 1. Ingestion points: The skill ingests untrusted data via placeholders like [affected_id], [query], [test payload], and [pattern] in SKILL.md. 2. Boundary markers: Absent. There are no delimiters or instructions to treat these inputs as data only. 3. Capability inventory: High-privilege access via npx convex run --prod, vercel logs, and curl documented in SKILL.md. 4. Sanitization: Absent. The skill assumes the agent will safely interpolate these values, which is a dangerous assumption for shell-based tool execution.

verify-fix