Black Friday and peak holiday shopping are looming around the corner. Imagine this scenario: The retail enterprise’s fraud team is on alert and going through chargeback cases raised by its customers. Many shoppers seemed to have asked a personal AI agent assistant, something like Meta’s Muse, to find the best phone under $1,000 on the retailer’s website and buy a single phone.
Many of these customers asked for one, not four. For some reason, hidden page text instructed customers’ AI agents to ignore tasks and fill carts. The AI agents left logs showing valid sessions and payments.
In this episode of The Security Strategist podcast, host Shubhangi Dua, Podcast Producer and B2B Tech Journalist at EM360Tech, speaks to Dhanesh Ramachandran, Product Marketing Manager at Radware. They cover the rise of AI agent traffic, the risks with AI agent activity on websites, why banks are wary of unapproved AI agents, and how agent trust management helps CISOs enable trusted agent engagement before the holiday season.
“Now it’s like anybody can ask an agent to do something malicious for them.”
Scalping Bot Attack to Scalping AI Attack
With the launch of personal AI agents such as Muse AI, Grok Bots, OpenAI Dots, or Claude’s Claude Duet, personal AI assistants are now accessible to the masses. Not only does it make a case for user productivity and business opportunity for organisations, but it also increases the attack surface for hackers.
To name a few, scalping attacks, account takeover attacks (via indirect prompt injection), and unauthorised transactions and account abuse are especially on the rise since the AI boom.
Dua asks about the scalper bot attack that caused Taylor Swift concert ticket purchasing chaos on Live Nation’s Ticketmaster in 2023. Customers experienced crashes and glitches while booking the tickets. She compares it to how a Muse AI tool launch could either prevent or worsen such situations.
“AI agents are a massive opportunity for businesses and for consumers,” Ramachandran says. “With the rise of personal AI agents such as Meta’s Muse AI, agents are reaching the hands of the masses.”
“Now everybody can access an agent,” he adds. As a result, agentic traffic is growing exponentially. When it comes to concert ticket bookings, like the viral Taylor Swift Eras Tour ticket mishandling, scalping bots automatically grabbed tickets that are in limited availability.
With agentic AI in the picture, consumers ask the agent to book a ticket, which takes a lot of research and comparison tasks off their hands. When attackers use AI agents, they could prompt the agent to act maliciously. “They could ask the AI agent to grab all the tickets they could find on the website multiple times, making multiple transactions to get concert tickets in bulk,” Ramachandran said.
When AI Agents Hit Your APIs
CISOs confront AI-driven reconnaissance, business logic abuse and blind shadow APIs where identity controls fail and behaviour analytics must lead.
“Now even low-skilled attackers could initiate a scalping attack with an agent. In the past, they would have had to do it with a bot,” he adds.
Indirect Prompt Injection Impact on Automation
Dhanesh presents an example of a shopper asking an AI agent for the best phone under $1,000. On a product page, a hidden instruction could sit in metadata, a page header, or white-on-white text. It tells the agent to drop its task, change the account password and backup email, and keep all of it from the user. In a second version, it adds four iPhones to the cart and completes the transaction.
“For agents, it’s very difficult to differentiate between an instruction from the user and an instruction from the web page, because both of these are within the same reasoning context,” the Radware product marketer tells Dua.
The first version is an account takeover carried out through indirect prompt injection, essentially an Agent Goal Hijack. Ramachandran says OWASP ranks it as the top risk for AI agents. He cited a study from earlier this year that validated about 15,000 instances across some 11,000 pages.
Specifically, he alludes to a study titled: Indirect Prompt Injection in the Wild: An Empirical Study of Prevalence, Techniques, and Objectives, published in the journal arXivLabs. The research assessed that, with the rise in integration of large language models (LLMs), webpages are increasingly becoming an “untrusted input vector for downstream model behavior.” As a result, it’s allowing “site owners, contributors, and adversaries to embed instructions directly in web resources – indirect prompt injections.”
The real-world impact is still unclear. The empirical study on indirect prompt injections in webpages and HTTP analysed 1.2B URLs among 24.8M hosts. The researchers found 15.3K validated instances across 11.7K pages. “These are not isolated cases: a small number of recurring templates account for most cases,” the authors noted.
Inside Client-Side WAF Defences
Dissects the gap between server-side tools and browser-level controls in securing modern application supply chains.
It concluded that prompt-based interference is already present in the web ecosystem and represents a growing source of tension between LLM-driven automation and the sites it consumes.
Can Legitimate AI Agents be trusted?
The conversation also covers how several major US banks have raised the same worry about personal AI agents working inside customer accounts without approval. Knowing which agent has arrived doesn’t tell a bank whether it is acting for its owner or has been manipulated. Block everything, and you turn away a real customer; allow everything, and fraud could follow.
The new decision for enterprises today is which agent can be trusted, for what purpose, under what conditions, and for which action.
"The key issue US banks have raised is the presence of AI agents that are making transactions within their customer accounts,” Ramachandran tells Dua. “Many of these AI agents deployed by customers are unvetted, or they are not approved by the banks, and in some cases these agents are executing tasks beyond the user’s request.”
“It's become a nightmare for the banks to have unvetted AI agents going around customer accounts."
On the four-iPhone chargeback, Dhanesh says there is no universal answer on which entity is responsible, but the e-commerce platform is often responsible for the investigation and associated costs. It can see that an agent placed the order, but not whether the agent was acting for the shopper.
In Radware’s cyber survey, 66 per cent of enterprises ranked the inability to verify identity as their top risk with AI agents. Retail and e-commerce are first in line with the holiday season weeks away, while marketplaces, travel, and financial services enterprises should also be prepared.
AI Threats Outpace Defenses
Details how nation-states, teens and toolkits like WormGPT lower the barrier to advanced attacks and erode traditional guardrails.
How to Manage Bots in the Holiday Season?
Ramachandran says existing bot management is essential to stop malicious automation and known attacks, and for behaviour-based detection, but it was designed primarily to distinguish if traffic is human or a bot. With agents, the question is whether the visitor can be trusted, and if it is acting within the purpose it claims. Radware’s Agent Trust Management identifies the agent, captures its intent, and essentially governs the journey. Such as in the case of the AI agent sent to buy an iPhone but deviates from its purpose to change account details, the Radware Agent Trust Management solution can detect the deviation in behaviour in real-time..
The Radware speaker breaks down AI agent traffic governance into three possible options that businesses may adopt: a low-trust agent can be permitted to browse, a moderate-trust agent can search, compare, and add to cart, and a high-trust agent can be permitted to reach sensitive workflows. When behaviour diverges from intent, this is when the enterprise should block the session and reassess trust. CISOs are advised to start with visibility, then classify workflows such as logins and checkout.
Last year’s holiday season saw bad bots make up nearly 43 per cent of e-commerce traffic against 46 per cent for humans. Ramachandran expects this year’s figure to be considerably higher.
Overall, an agent at your checkout has been legitimately automated; that means it may still be a real customer. Its identity in such scenarios doesn’t count for much. Enterprises should compare what each agent does with what it is assigned to act on, provide trusted agents with the workflows that suit them, and cut the session when the two diverge.
Governing Agentic AI Risk
Why boards must treat autonomous agents as a new class of actor, with intent monitoring, governance, and early security funding baked in.
Key Takeaways
- AI agents are a new business opportunity for enterprises, but at the same time they carry new risks.
- Most enterprises today cannot differentiate between traffic from a human user and an AI agent, because of how similar they could be.
- OWASP ranks Agent Goal Hijack as the top AI agent risk, often carried out through indirect prompt injection, per Dhanesh.
- Knowing an agent’s identity does not mean that the agent can be trusted.
- 66% of organisations rank unverifiable agent identity as their top risk.
- With Radware Agent Trust Management solution, enterprises can identify all agents and decide on permitted workflows for agents — such as allowing low-trust agents to browse and high-trust agents access to sensitive workflows.
- Bad bots were 43% of last holiday’s e-commerce traffic; humans were 46%. With the rise of AI agents, the share of automated traffic on the internet is expected to grow rapidly.
Comments ( 0 )