DruAI — agentic threat model

Built on Amazon Bedrock AgentCore, utilizing foundational models hosted by AWS. Primary threats include adversarial prompt injection to bypass safety filters and model alignment issues that could lead to incorrect risk assessments or destructive recovery actions.

Interacts directly with Druva's data security platform and telemetry analysis. Threats include telemetry data poisoning to mask malicious activities and unauthorized data exfiltration via RAG queries over sensitive enterprise backups.

Uses Amazon Bedrock AgentCore to orchestrate Data, Help, and Action Agents. Threats involve insecure tool integration and tool misuse, where the Action Agent could be tricked into executing unauthorized remediation or destructive recovery workflows.

Not certain from the listing — while it is embedded in Druva's enterprise platform and likely hosted securely on AWS, the specific sandboxing, network isolation, and secrets management policies for the agent execution environment are not disclosed.

Not certain from the listing — although the agent performs telemetry analysis on the platform, the specific observability, logging, and guardrails applied to the LLM inputs/outputs themselves are not detailed.

Not certain from the listing — enterprise compliance (e.g., SOC2, HIPAA) is implied by Druva's core business, but the specific access controls, authorization boundaries, and audit trails governing DruAI's actions are not explicitly defined.

Features a multi-agent ecosystem consisting of Data, Help, and Action Agents. This introduces threats of agent-to-agent trust abuse, where a compromised Help Agent could feed malicious instructions to an Action Agent, leading to cascading failures in recovery workflows.