Relevance AI — agentic threat model

Supports multiple foundation models (multi-model support), exposing the platform to model-specific vulnerabilities such as adversarial prompt injection, model alignment drift, and potential data leakage depending on the underlying third-party LLM providers used.

Not certain from the listing — while the platform supports custom tools and integrations, specific vector database configurations, RAG pipelines, or data ingestion security controls are not detailed.

High risk of tool misuse and insecure tool integration due to the 'tool builder for integrations' feature, which allows users to construct custom agent skills. Vulnerabilities in orchestration code could lead to unauthorized tool execution or prompt injection-driven tool hijacking.

Not certain from the listing — details on hosting environments, sandboxing of custom-built tools, and network isolation for multi-agent execution are not specified, though SOC 2 Type II compliance suggests standard enterprise infrastructure controls are in place.

Not certain from the listing — specific evaluation frameworks, real-time guardrails, or observability tools for monitoring multi-agent interactions are not described in the directory listing.

Demonstrates strong compliance posture by explicitly achieving SOC 2 Type II compliance, indicating established security controls, audit trails, and organizational security policies to protect tenant data.

High exposure to agent-to-agent (A2A) trust abuse and cascading failures due to its core feature of deploying multi-agent systems. A compromise in one agent or tool could laterally propagate across the entire deployed AI workforce.