building-chat-interfaces
Fail
Audited by Gen Agent Trust Hub on Feb 16, 2026
Risk Level: HIGHCREDENTIALS_UNSAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION] (HIGH): High-risk indirect prompt injection surface identified in
SKILL.mdandreferences/chatkit-integration-patterns.md. The skill implements features that scrape untrusted content from web pages (headings, meta descriptions) and user-selected text, which are then directly interpolated into the agent's system prompt using f-strings without sanitization, escaping, or boundary markers. This allows external data to override agent instructions. - [CREDENTIALS_UNSAFE] (HIGH): The skill explicitly instructs the AI agent in
SKILL.mdto include long-lived user credentials (access_token) in tool call arguments. This pattern creates a high risk of credential exfiltration if tools are malicious or if the agent is subverted by the aforementioned prompt injection surface. - [EXTERNAL_DOWNLOADS] (LOW): Frontend integration in
SKILL.mdandreferences/nextjs-httponly-proxy.mdrelies on remote script loading fromhttps://cdn.platform.openai.com. Although the source is within the trusted scope for OpenAI, runtime execution of remote code remains a security consideration.
Recommendations
- AI detected serious security threats
Audit Metadata