As enterprises accelerate the deployment of AI proxies, verifying whether these systems are secure and controllable has become a practical issue in the procurement process. AIUC, founded by a former employee of Anthropic named Rune Kvist and the Chief Operating Officer of a former AI security research institution named METR Rajiv Dattani, is attempting to turn this process into a standardized service.
Announces completion of a $40 million Series A financing round
AIUC is officially known as Artificial Intelligence Underwriting Company. The company announced on Tuesday that it had completed a $40 million Series A financing round, led by Ribbit Capital, with First Harmonic also participating in the investment.
Prior to this, AIUC had completed a seed round of financing of $15 million, with investors including the funds NFDG, Emergence, and Terrain under Nat Friedman, as well as Ben Mann, one of the co-founders of Anthropic. To date, the company has raised a total of $55 million in financing.
AIUC indicates that the current customers include Cursor, Lovable, Harvey, and ElevenLabs.
Mainly providing AI agency for third-party audits
The core product of this company is to establish a third-party audit and certification layer for AI agents. Its approach is similar to the common compliance verification models in the field of network security, which involve setting unified standards first and then using external testing to determine whether the system meets those requirements.
AIUC Referring to widely used cybersecurity standards, SOC, the company has introduced a standard named AIUC-1 along with accompanying testing services. The company states that many banks, hospitals, government agencies, and military organizations do not refuse to deploy AI due to insufficient model capabilities, but rather because there is a lack of verifiable security guarantees at this stage.
To establish standards, AIUC organized approximately 250 safety and risk managers to participate in the discussion. These individuals are also decision-makers when it comes to corporate procurement through AI agents. Based on this, the company identified the issues that purchasers are most concerned about and then transformed these requirements into testing projects.
Approximately 5,000 tests cover jailbreaks and leaks.
In terms of actual implementation, AIUC will have AI act as a proxy to go through approximately 5,000 tests, with a focus on observing its performance in scenarios such as jailbreaking, hallucinations, and data leakage. After the tests are completed, the system will generate a report of about 100 pages, listing which aspects of the proxy's performance are stable and which still pose risks.
It is worth noting that AIUC also uses AI proxies to execute some tests and utilizes AI to analyze the test data, but the final audit results are still subject to manual verification.
This mode has similarities with some of the work of METR. In the past, METR mainly provided independent testing for cutting-edge model laboratories, with a focus on performance evaluation, that is, whether the agents could stably complete tasks. When investigating Hugging Face-related incidents before, METR was also used to participate in independent research.
Anthropic The CEO Dario Amodei has also recently called for a slowdown in the development pace of cutting-edge technologies AI, and mentioned the possibility of introducing third-party evaluation agencies to observe and verify security performance. AIUC's approach is not to directly embed solutions at the customer site, but to provide companies with an independent external assessment to help them determine the scope of availability and risk points before procurement AI.
