Harbor raises $10 million to protect businesses from AI agent risks

US-based startup Harbor has secured $10 million in seed funding, led by Base10 Partners, Abstract, and 8VC. The company aims to combine an AI agent risk assessment platform with insurance coverage for financial losses that occur when preventive controls fail.

From credit to refunds: why delegating decisions to AI is risky

An AI agent can approve loans, process refunds, or execute debt collection. When these decisions go wrong, the company that implemented the system bears the loss. Harbor fills this gap by providing a protective layer between task delegation and financial exposure.

Risk assessment without a claims history

Traditional insurance pricing relies on data about the frequency and severity of losses. However, AI agents lack a comparable claims history to older activities. Instead of relying on past records, Harbor directly analyzes the agent’s performance.

Customers provide two types of information: the model architecture (tools, data, and permissions) and records of previous decisions and actions taken during operation. The company uses a 45-risk framework, developed in partnership with institutions like the Stanford AI Safety Lab. For each application, relevant risks are mapped and divided into possible failure modes.

The analysis considers the frequency of errors observed in tests and production, the potential cost of each failure, and variables such as permissions, transaction limits, and industry. The results are compared with an internal test database and public data to determine the insurance premium.

When the model, tools, or data are updated, Harbor re-evaluates the agent and adjusts the price based on the new exposure. “We measure the agent directly. You can’t do that with a human employee, but you can with an agent,” explains Jesús Santos, co-founder of the company.

Real-time intervention before execution

The platform also inserts a control point between the agent’s decision and its execution. In less than 100 milliseconds, the system compares the proposed action with the company’s rules, allowing it to block the action, permit it to continue, or forward it to a human responsible.

For example, a bank that allows automatic refunds but requires human approval for amounts above $5,000 may have its policy violated if the agent attempts to refund $7,200. The platform identifies the violation before payment, blocks the transaction, and generates a justification for review.

In the context of Pix, where transfers are completed in seconds, the interval between decision and action is critical for preventing errors from being repeated thousands of times.

What the insurance covers and who assumes the risk

The policy offered by Harbor includes six categories of failure: incorrect responses or hallucinations, unauthorized actions, bias and discrimination, breach of contracts, data leakage, and intellectual property or defamation violations.

The coverage includes direct losses of the company, civil liability to third parties, defense costs in regulatory proceedings, and incident response. Harbor handles pricing and underwriting, while partner insurers provide financial capacity and assume the risk.

“That’s the purpose of insurance: it needs to pay when prevention fails,” says Santos, explaining that the protection remains valid even if the startup’s internal controls do not work.

Negotiations with major Latin American banks

Latin America is one of Harbor’s three strategic regions, alongside the United States and Europe. The company is in talks with three of the five largest Latin American banks, whose names have not been disclosed. The institutions’ primary interest is in real-time control before an action is executed.

Among the tasks banks want to automate are credit approval, limit increases, debt collection, payments, transfers, and compensation decisions. The challenge, according to Santos, is to explain to boards, regulators, and insurers how to act when the agent makes a mistake.

Frequently Asked Questions

• What does Harbor do? It offers a platform to assess AI agent risks, intervene in decisions before execution, and structure insurance for covered losses.
• How much money did the company raise? $10 million in a seed round with Base10 Partners, Abstract, and 8VC.
• How does Harbor calculate the insurance price? It analyzes the agent’s architecture and records, estimates the frequency and cost of failures, and compares the results with its test database.
• Who assumes the financial risk of the policies? Harbor handles pricing and underwriting; partner insurers assume the financial risk.
• Does Harbor already have customers in Brazil? The company is negotiating with three of the five largest banks in Latin America but has not confirmed paying customers in the country.