Dive Brief:
- The American Institute of CPAs and Artificial Intelligence Underwriting Company are rolling out a joint assurance certification program that offers an independent review of business risk related to AI systems and agents, they said Monday.
- The initiative combines a System and Organization Controls (SOC 2) report with AIUC-1 certification, the organizations said, describing it as the first comprehensive, independent certification program for AI agents.
- “To date, the primary focus of AI assurance has been governance and compliance, which is important, but it isn’t enough to just have policies in place — what is critical is to demonstrate that those policies are being followed and that AI systems and agents are secure, safe and operating reliably, as evidenced by testing and independent third-party assurance,” Amy Pawlicki, AICPA’s vice president of assurance and advisory innovation, said in an email.
Dive Insight:
The move comes amid growing concerns about the risks posed by AI systems capable of acting on their own. Disclosures of unintended AI-agent behavior have intensified debate over the pace of development, with some industry leaders calling for a slowdown.
Lawmakers, meanwhile, are proposing new liability rules to address potential AI dangers. A draft bill released earlier this month by Rep. Lori Trahan, D-Mass., would allow private parties to sue developers when an AI system injures a third party through conduct that would constitute negligence, an intentional tort or a crime if committed by a person.
Another measure, introduced by Sens. Josh Hawley, R-Mo., and Chris Murphy, D-Conn., would hold AI-agent operators criminally and civilly liable under the Computer Fraud and Abuse Act for certain hacking-related harms.
“As legislators and regulators are increasingly concerned about the risks of AI, it is critical that developers and deployers of AI systems and tools have the ability to demonstrate in a meaningful way how they are mitigating this risk — this joint offering supports this,” Pawlicki said.
As part of the certification program, SOC 2 examines whether an organization’s controls are appropriately designed and operating effectively, while AIUC-1 tests the behavior of AI agents, according to the organizations.
“Artificial intelligence is a transformative technology for many organizations, but misfires have raised serious concerns about its safety and security in critical business operations,” they said in a Sunday release.
AIUC-1 testing targets risks such as prompt injection, in which malicious instructions hidden in emails or documents can manipulate an AI agent, as well as data leakage and actions that exceed an agent’s authority. An organization can have strong access controls and still face risks if an agent misuses access it legitimately holds, AIUC said in emailed comments.
Technical testing evaluates agents against large numbers of scenarios tailored to their intended work. For example, a finance agent could be tested against manipulated invoices or payment instructions. Other tests could examine whether an agent follows malicious instructions in an uploaded file, exposes internal data or invents financial figures or company policies.
When testing identifies weaknesses, the company can address them and have the agent tested again to verify the fixes, AIUC said. Results from both rounds are included in the certification report.
AIUC-1 also requires quarterly technical retesting to account for changes in AI models, tools and prompts. SOC 2 reports and AIUC-1 certifications require annual renewal, according to the announcement.
The initiative is aimed at technology companies seeking to demonstrate the reliability and security of their AI systems to businesses that use them, Pawlicki said. For companies evaluating AI vendors, the certification could provide additional evidence to consider alongside a SOC 2 report when assessing the risks of relying on third-party systems.
Pawlicki said using the certification does not completely eliminate risk. “Ultimately, there are risks with any technology,” she said.
Pilots are under way that will inform a program launch targeted for the first quarter of 2027, Pawlicki said.
