Harvey AI — agentic threat model

Utilizes proprietary foundation models customized and trained on legal datasets. Primary threats include model stealing of proprietary legal fine-tunes, data poisoning of custom training sets, and mis-aligned outputs such as hallucinated case law or legal precedents.

Processes highly sensitive legal documents, contracts, and litigation data. Key threats include data exfiltration of privileged client information, knowledge-base poisoning of legal reference materials, and unauthorized access to vector stores containing proprietary case strategies.

Not certain from the listing — details on the specific orchestration framework, planning mechanisms, or tool-calling protocols are not disclosed. However, insecure tool integration for document parsing and legal research queries could lead to prompt injection exploiting underlying parsers.

Not certain from the listing — hosting environment, sandboxing for document parsing, and secrets management details are not provided. Insecure document ingestion pipelines could expose the infrastructure to remote code execution via malicious PDFs or Word documents.

Not certain from the listing — no explicit details are provided regarding real-time guardrails, drift detection, or evaluation frameworks for legal accuracy. The lack of transparent observability could lead to undetected drift in legal reasoning or compliance advice.

Not certain from the listing — while the tool assists with regulatory compliance, the platform's own internal security controls, identity management, and audit logging are not specified. Compliance with legal industry standards (e.g., SOC 2, HIPAA, attorney-client privilege boundaries) is critical but unverified from the listing.

Not certain from the listing — there is no mention of multi-agent coordination, external agent marketplaces, or agent-to-agent trust boundaries. The ecosystem appears limited to a single-platform assistant model.