Dust AI — agentic threat model

Not certain from the listing — The specific foundation models powering Dust AI are not disclosed. Standard LLM threats like prompt injection, adversarial reprogramming, and misaligned outputs remain highly relevant as they could hijack the agent's reasoning chains.

Dust AI integrates deeply with company knowledge sources and internal tools. This creates significant exposure to data/knowledge-base poisoning (where malicious internal documents alter agent behavior) and unauthorized data exfiltration via tool connectors.

The platform supports multi-step workflows, reasoning chains, memory, and tool connectors. This introduces severe risks of tool misuse, insecure tool integration, and memory poisoning, where malicious inputs persist in the agent's state and compromise subsequent steps.

Not certain from the listing — The hosting infrastructure, network sandboxing, and secrets management for API integrations are not detailed, leaving potential risks of container compromise or lateral movement unverified.

Dust AI explicitly features 'observability' as a key capability. While this helps mitigate blind spots, robust guardrails and anomaly detection are required to prevent evaluation gaming or undetected drift in complex reasoning chains.

Not certain from the listing — Although described as 'Enterprise Software' that 'securely manages data', specific compliance certifications (such as SOC2, ISO 27001) or granular role-based access controls (RBAC) are not explicitly detailed.

Not certain from the listing — While teams can build multiple custom agents, the listing does not specify if these agents interact directly with each other or if there is an active multi-agent ecosystem/marketplace, which would introduce risks of cascading failures or agent-to-agent trust abuse.