While safeguarding autonomous agents from fraud comes with its unique challenges, time, and experience, it is on the good side, Accertify’s vice president of artificial intelligence and machine learning, Conrad Kennington, said. With 20 years of experience in e-commerce and fraud prevention, Kennington knows how to adapt to new technologies.
Autonomous agents are the next stage of e-commerce, and share similarities with past developments. The Internet allowed us to shop brands from home – we could shop multiple stores from our living room. Amazon correlates the stores on one site. Now, we don’t even have to shop – autonomous agents shop for us.
Protecting autonomous agent technology requires different tools
Online interactions leave clues like browser data, device signals, and form fields, but every new stage eliminates some of these traditional fraud signals. Autonomous agents could also foster friendly fraud, where consumers forget they programmed an agent. If single-use tokens are used, behavioral context is erased, leaving fraud prevention teams little more than a shipping address.
“Agentic shoppers will be attractive to fraudsters for that reason,” Kennington said. “They’re basically an unlimited number of virtual hitmen who are doing your bidding. You don’t even need how to code; the barrier to entry couldn’t be lower.”
Kennington likens autonomous agents to a sheriff leaving town and leaving instructions with a deputy. Do those notes cover every conceivable case? While it’s a tall order, experience has Kennington confident that fraud prevention providers will adjust. The signals are still there; firms must discover new ones and adapt familiar ones to new technologies.
“Signal loss is not new,” he said. “With each mobile device and browser, data points are lost.”
“I understand. People want to be anonymous online. They’re tired of breaches and having their data stolen.”
I understand. People want to be anonymous online. They’re tired of breaches and having their data stolen
That’s the reality, so the industry has to continually adapt. Work with what you know, such as analyzing behavior patterns. If someone commits fraud on a device once, they’re likely doing it again. Develop ways to fingerprint devices.
Enrich merchant and network data. Leverage BIN data.
“Each signal lost helps determine another one,” Kennington said. “How must we engineer features and change approaches? Fraudsters become more sophisticated, and so do we.”
Agentic shoppers are a (very) small part of autonomous agents
Large language models, around since the late teens, began attracting attention in late 2022 when ChatGPT figured out how to scale them. Kennington said the industry has had four years to observe, learn, and strategize.
He cautions that autonomous agents are a very small piece of a global automation phenomenon. Companies seek to automate everything, and autonomous agents don’t even crack the top 10; Kennington said less than 1% of agents are shopping.
“We see adoption for LLMs as very rapid, but I believe adoption for agentic shopping will be more lagging and lower, because there is still a fundamental lack of trust in LLMs to do things that might affect your pocketbook,” Kennington said.
That buys the industry some time, as consumers won’t trust autonomous agents until appropriate protocols are in place. Still, fraud prevention firms are working to be ready when adoption increases. Adoption has so far been mostly limited to tech-savvy types and younger consumers.
Kennington believes autonomous agents will initially be used for essentials, or perhaps gifting, in which a shopper describes the intended recipient, the agent makes suggestions, and the purchase is completed. It won’t be used for luxury goods or clothing anytime soon, which shoppers want to see for themselves. Commodities and travel could be prime early-adoption sectors.
Key considerations when safeguarding autonomous agents
Identity certification is key to safeguarding autonomous agents.
“That’s why we’re getting in early because resolving an identity is important for stopping fraud,” Kennington said. “If you can’t resolve who’s making the purchase, the fraudster can just go do everything on your behalf.”
Agentic shopping makes that challenging. Someone (or something) might have access via your username and password. That doesn’t mean they have authorization.
“Those same concepts can be transferred to an agent by saying you might be authorized to make a purchase, but is it authenticated to the person who owns that payment token?” Kennington asked. “That is something that is solvable, and we’re trying to solve in the protocol.”
Digital signatures are important parts of proving identity. Kennington said they’ll ensure the source buyer’s identity is legitimate and cannot be tampered with. Shoppers using autonomous agents must also provide clear, tamper-proof instructions.
Other issues include how to prepare autonomous agents to handle follow-up questions about sites like travel insurance or other add-ons.
“Having an agent shopper know all of those things in advance is a lot of extra effort where people may just do it themselves,” Kennington suggested.
While there are remaining kinks to be worked out, Kennington is confident fraud prevention firms like Accertify will succeed in protecting autonomous agents and agentic commerce.
“It’s going to be years, not months. We’ve seen early adoption in niche sets, and the race is on. The long-term challenge is going to be how do we make it safe.”