Ali Ikhwan
Introduction: The New Frontier of Fraud
On October 8, 2026, Riskified (NYSE: RSKD) officially unveiled Agent Identity Risk Intelligence, a sophisticated suite of tools designed to tackle the next generation of digital commerce threats. As AI personal assistants—such as Meta’s Muse, various GPT-based agents, and emerging autonomous shopping bots—begin to represent consumers in everyday transactions, the traditional pillars of fraud detection are crumbling.
Riskified’s solution seeks to identify the specific AI agent behind an order or a customer service request, resolve that digital footprint to the actual human consumer, and provide a real-time risk signal to merchants. While the technology is initially debuting on the Shopify platform, it is destined for integration across Riskified’s existing APIs, the AI Agent Approve MCP server on AWS Marketplace, Decision Studio, and Zendesk. With a limited beta scheduled for December 2026 and a full general availability (GA) rollout expected in 2027, the market is bracing for a fundamental change in how identity is verified.
Yet, despite the technical sophistication of the launch, the company’s internal framing offers the most sobering assessment of the landscape: "Authenticated is not the same as trusted."
The Three Risk Questions: Redefining Security
To understand the significance of Riskified’s move, one must analyze the "three risk questions" that define the modern transactional environment. As the digital economy pivots toward machine-to-machine interactions, the definition of a "legitimate" user has become increasingly murky.
1. The Identity Question (Who is the agent?)
This is the baseline requirement. Merchants need to know if the entity interacting with their storefront is a human, a bot, or an AI agent acting on behalf of a human. If it is an AI, which model is it, and what are its credentials?
2. The Authorization Question (What is the agent allowed to do?)
This is the territory Riskified is staking out with its new intelligence suite. It addresses the permissions granted to the AI. Is this agent authorized to make a purchase? Is it authorized to initiate a return? Just because an agent has the correct OAuth tokens or API keys does not mean it is acting in the best interest of the merchant or even the user.
3. The Intent Question (What is the agent actually doing?)
This is the gap that remains largely unaddressed by standard identity checks. An agent might be "authorized" by a stolen account or a malicious "refund-as-a-service" (RaaS) ring. When a malicious actor uses a legitimate identity to grant permissions to a rogue AI, the result is fraud executed at machine speed.
Riskified’s release is unusually candid about this danger. "An agent authorized by a stolen account is still an account takeover. An agent authorized by a refund-as-a-service ring, whose members hold real accounts and grant real permissions, is still fraud."
Chronology of a Shifting Landscape
The necessity for this toolset was underscored by the landmark legal developments of early 2026.
- April 2026: A U.S. federal court sentenced eleven co-conspirators involved in the "Artemis Refund Group." This syndicate operated refund fraud as a subscription-based service, systematically targeting retail giants like Amazon, Walmart, Target, and Wayfair. The case served as a wake-up call to the industry: the defendants utilized real accounts and possessed real permissions. The fraud did not lie in the identity—it lay in the claims.
- October 8, 2026: Riskified officially announces the Agent Identity Risk Intelligence platform, acknowledging that the traditional "verified identity" is no longer a sufficient proxy for safety.
- October 11, 2026: Initial live tests performed by ScriptMasterLabs (SML) reveal the inherent weaknesses in current heuristic-based decision gates.
- December 2026: Anticipated start of the limited beta phase.
- 2027: Projected full General Availability (GA) of the product suite.
Supporting Data: The Failure of Legacy Heuristics
A live test conducted by ScriptMasterLabs on October 11, 2026, highlights the dangerous asymmetry in current risk assessment models. Using a decision gate where scores ≥0.80 result in auto-approval, 0.50–0.79 require manual confirmation, and <0.50 trigger escalation, the researchers tested two distinct scenarios.
The test demonstrated that a "blunt heuristic" treats a refund-as-a-service ring—hiding behind a verified, high-trust identity—identically to a loyal shopper requesting a routine return.
The Asymmetry of Approval
Riskified’s current Decision Studio often auto-approves "low-risk agent orders from verified identities." Under this logic, both a legitimate consumer’s AI assistant and a malicious bot operating within a RaaS ring would be cleared for entry. The identity check validates the credentials, but it fails to evaluate the instructions given to the AI.
As long as merchants rely on identity alone, they remain vulnerable to "authorized" fraud. The data suggests that merchants must move toward a model where identity risk and instruction judgment are two distinct, yet integrated, instruments. Failing to wire both into the same execution path effectively leaves the front door open to sophisticated, high-speed automated attacks.
Official Responses and Strategic Positioning
Riskified’s leadership has positioned this product not as a replacement for existing fraud filters, but as an essential evolutionary layer. By integrating with AWS Marketplace and Zendesk, they are attempting to move the point of decision-making closer to the source of the interaction.
Industry analysts note that Riskified is attempting to commoditize "instructional integrity." By scoring the intent behind the AI agent, they are effectively creating a new category of "Agent-Aware" security. The company’s focus on the AWS Marketplace is particularly telling; by making the AI Agent Approve MCP server available in a standard cloud environment, they are lowering the barrier to entry for merchants who are currently struggling to manage the influx of AI traffic on their platforms.
Implications: The Future of Autonomous Commerce
The implications for the retail sector are profound. As AI agents become ubiquitous, the "conversion" metric—long the holy grail of eCommerce—will be redefined.
1. The Death of Static Trust
Merchants can no longer rely on static trust scores associated with email addresses or credit cards. Trust must now be dynamic and session-specific, evaluated based on the behavior and the specific instructions being processed by the AI agent.
2. The Rise of "Instructional Forensics"
As demonstrated by the Artemis Refund Group, the future of fraud is not in identity theft, but in "permission abuse." We are entering an era of "Instructional Forensics," where merchants must analyze the prompt engineering or the underlying logic of the agent making the purchase or request.
3. The Platform Bottleneck
By anchoring their solution in the Shopify ecosystem and extending it through AWS, Riskified is creating a "choke point" for commerce. While this increases security, it also centralizes a significant amount of power over what constitutes a "valid" transaction. Merchants will need to carefully consider whether their risk appetite aligns with the scoring models provided by a single third-party vendor.
4. Regulatory Pressure
The shift toward automated decision-making in retail will inevitably draw the attention of regulators. If an AI agent is denied a transaction based on an "Agent Identity Risk" score, the transparency of that decision-making process will become a legal concern. How does an agent appeal a rejection? Who is liable when an AI is incorrectly identified as a fraudster?
Conclusion
The launch of Riskified’s Agent Identity Risk Intelligence is a recognition that we have reached a saturation point in identity-based security. The industry has spent years perfecting the art of knowing who is behind the screen. Now, it must learn to decipher what they are doing and, more importantly, why they are doing it.
As we move toward the 2027 GA rollout, the merchant who succeeds will be the one who realizes that their security posture is only as strong as their ability to distinguish between a loyal customer’s digital assistant and a high-speed, automated fraud ring. In the age of the AI agent, trust is no longer a given—it is a variable that must be calculated in milliseconds.
For further technical details, including the specific curl commands to score agent instructions, merchants and developers are encouraged to visit the official documentation provided by ScriptMasterLabs.
Muslim
Ammar Sabilarrohman
Lina Irawan
Laily UPN
Nana