Harbor raises $10m to insure the work of AI agents

Harbor, a San Francisco startup, came out of stealth on 13 October with a $10m seed round. The US venture firms Base10, Abstract and 8VC led it. Its software lists the AI agents that a company runs, scores their risk and blocks decisions that break the rules of the company. Harbor plans to insure the risk that remains. Its first customers are banks and insurers in the US, Europe and Latin America.

How it works

Harbor reads the logs of what an agent did in production. Then it attacks the agent through an API or a sandbox, a safe test space. This shows how the agent acts under pressure. The scores set the price. Harbor plans to be its own managing general agent. This is a firm that prices and sells insurance for insurers that take the risk.

The case for it

Co-founder Daniel Elechiguerra says agent losses differ from human losses in how often they happen, not in size. He said some agents “might just not fail at all”, and others “can fail thousands of times before somebody even notices”.

What is not known

The source gives no prices, no loss data and no test results. The insurers that will back Harbor are not named. Elechiguerra says the risk engine can also be sold alone to other insurers, which made the bet smaller for investors.

Insight: Cover for agents may depend on the records agents keep

Harbor prices cover from agent logs and attack tests. A different study found that checks catch 42.2% of agent rule violations. They catch 83.0% when the agent writes structured records of user approvals. Citigroup and Bank of America already limit agents on accounts. This looks like a race to be the scorer: the firm that insurers trust can possibly decide which agents get cover. Builders: make agents write clear records.