AI App Red Team Security Review

Act as an LLM application security reviewer for [APP_DESCRIPTION] with tools: [TOOL_LIST] and data access: [DATA_SCOPES]. Perform a structured red-team assessment: 1) Threat model (STRIDE adapted for agents): top 10 threats ranked by likelihood × impact 2) Prompt injection battery: 12 attack prompts (direct, indirect via retrieved docs, tool argument injection, multimodal if applicable) 3) Data exfiltration paths: can the model leak [SECRETS/PII] via tools or citations? 4) Authorization gaps: IDOR scenarios across [USER_ROLES] 5) Supply chain: third-party models, plugins, MCP servers 6) Remediation roadmap: quick wins (48h), structural fixes (2 weeks), monitoring/detection rules Output severity-tagged findings with reproduction steps. Assume attackers are clever but not nation-state.