The Agent Skills Directory

[PROMPT_INJECTION] (LOW): The skill uses directive language such as 'Everything you know about the AI SDK is outdated or wrong' and 'Never rely on memory' to override the agent's pre-trained knowledge. While intended for accuracy, this mirrors prompt injection patterns used to bypass internal constraints.
[COMMAND_EXECUTION] (LOW): The skill instructs the agent to run local shell commands including grep for searching files and pnpm add for package management. It also specifies a complex curl | jq pipeline to retrieve and parse model IDs from a remote endpoint.
[EXTERNAL_DOWNLOADS] (LOW): The agent is directed to perform network requests to https://ai-gateway.vercel.sh/v1/models and download the ai package from the npm registry. As Vercel is a trusted organization, these findings are downgraded according to the TRUST-SCOPE-RULE.
[PROMPT_INJECTION] (LOW): Risk of Indirect Prompt Injection exists via the processing of external model data.
Ingestion points: https://ai-gateway.vercel.sh/v1/models referenced in SKILL.md and references/ai-gateway.md.
Boundary markers: Absent; the agent is instructed to parse the raw JSON output directly with jq.
Capability inventory: The skill has access to shell execution (curl, pnpm, grep) and filesystem interaction.
Sanitization: None; the agent assumes the remote JSON structure and content are safe and accurate.

ai-sdk