Harbor raises $10M seed to measure and insure AI agent risk

Why continuous measurement

According to Elechiguerra, companies already leave real decisions to agents, such as approving loans, pricing policies and moving money. When one fails, the company is liable, and insurers are starting to exclude that risk. One-off audits before deployment fall short because models, tools and data change constantly. He cited a 2024 Canadian ruling that forced Air Canada to honor a refund policy invented by its chatbot.

Harbor first inventories a company's agents, then scores each on three fronts: whether it does its job well, whether it can be tricked or leaks data, and whether it complies with applicable rules. It then checks each decision against the client's rules and blocks non-compliant ones before they run. The framework, built with Stanford and other partners, aligns with the NIST AI RMF, SR 26-2 and the EU AI Act, and measurements are repeated whenever something changes.

Market and plans

Chief of Staff Juan Vecino said the drivers differ by region. In the US, insurers can add standard generative-AI exclusions to general liability policies since January 2026. In Europe, regulation is the push, with the AI Act coming into force and the ECB asking supervised banks in July for a plan against AI-based threats.

After its launch drew coverage in about 70 outlets, Harbor chose to focus on regulated sectors such as banking and insurance. It already works with companies in Spain. The funding will go to the risk platform, more insurance capacity and growing the team in San Francisco and Spain. Elechiguerra expects that within three to five years no serious company will run an agent without continuous risk measurement and a way to transfer it.