Harbor emerges from stealth with $10M funding to handle ai-agent risk. October 13, 2026 New company Harbor said it has emerged from stealth to help with the risks of deploying AI agents. San Francisco-based Harbor starts with $10 million in seed funding from Base10 Partners, Abstract and 8VC. Harbor said it helps companies measure and mitigate the operational, security, and regulatory risks of AI agents, insuring them if the agent fails or causes a loss. Harbor works with banks, insurers and other regulated companies in the United States, Europe and Latin America. Historically, when an employee makes a mistake, the company is covered, said Harbor. Professional liability, cyber, and tech E&O policies pay for the damage. But with AI agents, a mistake is repeated - possibly thousands of times - until someone notices. When it comes to AI agents, insurers do not have the claims history to price the risk. So some insurers are excluding AI out of existing policies, Harbor said. Harbor is building the solution and works in two ways: * The first is the risk platform. Harbor finds every AI agent a company runs and builds an inventory of them. It measures the operational, security and regulatory risk of each one, using a risk framework built with research and enterprise partners and mapped to the main AI frameworks and regulations, from NIST AI RMF and SR 26-2 to the EU AI Act. It checks every decision against the company's own rules and blocks the ones that break them before they go through. Every time an agent's model, tools or data change, Harbor measures it again. * The second is insurance. The measurement is what makes the risk insurable. The risk that remains, Harbor insures. Insurance also changes the incentives: when the insurer pays for the loss, it has every reason to make the measurement rigorous and the controls work. "Companies need to understand the risk of their agents, and at the current pace of adoption and development, assessing the risk at a specific point in time seems useless. There is a need for an innovative solution for managing the risk in real time," said Jesús Santos, co-founder of Harbor. Santos is one of four founders who met at Stanford University. Daniel Elechiguerra, Cristian Bartolomé, and Luis Botín are the other founders of the company. Was this article valuable? Interested in agencies? Get automatic alerts for this topic.

Startup raises US$10 million to monitor AI agents and create insurance against failures. Harbor raised US$10 million to monitor AI agent risks, block failures and insure financial losses; the startup is negotiating with major banks in the region. Team from Harbor, the startup developing control and insurance tools for autonomous AI agents (Harbor/Disclosure) Artificial intelligence and technology reporter Published on October 13, 2026 at 9:59 a.m. An artificial intelligence (AI) agent can approve credit, refund money or execute a charge. If the decision is wrong, however, the loss can remain with the company that put the system to work. It is in this gap between delegating a task and transferring its financial risk that Harbor intends to build its business. "When an agent causes a loss, the company pays with resources from its own balance sheet," says Jesús Santos, co-founder of the startup, in an interview with EXAME. For him, the expansion of AI-related exclusions in traditional policies leaves companies exposed to errors that they previously associated with the actions of employees or conventional software. Harbor reports having raised US$10 million in an initial round with Base10 Partners, Abstract and 8VC. Headquartered in San Francisco, the company combines an assessment and control platform with insurance for the risks that remain after preventive measures. "We measure the risk of each agent, interrupt decisions that do not obey the company's rules and insure the risk that remains," says Santos. The focus is systems capable of executing actions autonomously, especially at banks, insurers and other companies in regulated sectors. The business follows the discussion about insurance for AI, which involves financial losses, discriminatory decisions and operational failures. Liability also depends on the contract. Santos divides the chain into three participants: the model provider, the party developing the agent and the company using it. In his assessment, the organization that puts the system into operation tends to assume the portion of risk that was not contractually transferred. When an external provider develops the agent, the executive says it is common to have a limitation of liability equivalent to 12 months of the contract. He notes, however, that this condition depends on the agreement signed. "The company using the agent retains all remaining liability," he says. In cases where development is internal, there is no external agent provider with whom to share this exposure. The concrete distribution of liability depends on the contracts and the rules applicable to each situation. Santos acknowledges that he does not know the conditions negotiated between model providers and their large clients. For Harbor, the commercial point is in identifying which losses remain with the contracting party and which can be covered by a specific policy. How to calculate a risk without decades of history. The price of insurance usually depends on estimates of the frequency and size of losses. AI agents, however, do not yet have a claims history comparable to that of older activities. Harbor says it faces this limitation by examining the system directly. "We do not wait for the claims history. We measure the agent directly. You cannot do this with a human employee, but you can with an agent," says Santos. The client provides access to two sets of information. The first describes the architecture: model, tools, data and permissions. The second gathers records of previous decisions and of the actions the agent takes during operation. According to the co-founder, the startup works with a framework of 45 risks, developed with research partners, including the Stanford AI Safety Lab. For each application, it selects the relevant risks and divides them into possible forms of failure. The analysis considers the frequency of errors, in testing and in operation, and how much each one can cost. Permissions, transaction limits, sector and purpose of the agent enter into this assessment. The results are compared with a database of partner tests and public data. According to Santos, this comparison guides the insurance price. "When the model, the tools or the data change, the risk also changes. That is why we measure the agent again after each change, and the price follows the new assessment," he says. "A test in June does not show how the agent operates in September." Between deciding to refund money and executing the payment. The platform can also act between the agent's decision and its execution. In that interval, it compares the intended action with the company's rules and allows, blocks or forwards the case to a person. Santos says Harbor's models make this assessment in less than 100 milliseconds. The timeframe is a measure reported by the company. To illustrate how it works, he describes a hypothetical example: a bank allows automatic refunds but requires human approval for amounts above US$5,000. The customer must also confirm their identity in the same session, and an account cannot receive more than two refunds in 30 days. If the agent decides to refund US$7,200, the platform identifies the violation of the first rule before payment. The action is blocked and the case goes to a person, accompanied by the justification. "In Brazil, Pix completes a transfer in seconds. Therefore, the interval between the decision and the action is the moment to interrupt the error," says Santos. When no rule covers the situation, Harbor says it compares the decision with the usual behavior of that agent and of similar systems. Amounts far above the standard, unprecedented actions or signs of manipulation can lead the case to human review. "The person decides and writes a new rule. Thus, the company's rules come from real decisions," he says. What the insurance is intended to cover. According to Santos, the policy developed by Harbor covers six categories of failure: hallucinations and errors in responses; incorrect or unauthorized actions; bias and discrimination; breach of contracts and service levels; data leaks; and violations of intellectual property and defamation. The proposal includes direct losses of the company, liability toward third parties, defense costs in regulatory proceedings and incident response, according to the contracted coverage. "Harbor does the pricing and underwriting. The insurers provide the capacity and assume the financial risk," he says. Underwriting is the assessment that determines acceptance of the risk and the insurance conditions. Asked about what happens if the startup's own controls fail, Santos says the protection remains valid. "That is the purpose of insurance: it must pay when prevention fails." The interview does not identify the partner insurers or detail deductibles, indemnity limits or exclusions. It also does not provide the price of the policies. Conversations with Latin American banks. Latin America is one of Harbor's three regions of operation, alongside the United States and Europe. According to Santos, the company is in conversations with three of the five largest Latin American banks. The names remain confidential. The executive does not confirm, in his answers, how many paying clients or agents in production the company already has in the region. He says the greatest interest in the negotiations is in real-time control, before an action is executed. Among the activities that the institutions want to delegate are credit approval, limit increases, debt collection, payments, transfers and decisions on insurance claims. In the founder's assessment, systems that work in tests may encounter difficulties in receiving authorization and entering real operation. The question is how to explain to the board, the regulator and the insurer what happens when the agent errs. "In each case, the decision has a direct financial result, and the agent makes it thousands of times a day," says Santos. "An error can repeat thousands of times before a person notices." Frequently asked questions. What does Harbor do? How much did the company raise? How does Harbor calculate the price of the insurance? Who assumes the financial risk of the policies? Does Harbor already have clients in Brazil? Follow everything about: Startups