During an authorized security assessment of a major insurance provider’s production website, an autonomous penetration testing agent discovered a critical-severity authentication bypass that could permit unauthorized account takeovers. The affected platform serves millions of policyholders and users.
The vulnerability was identified by an artificial intelligence agent developed by cybersecurity firm Reflectiz while it was systematically mapping the insurer’s authentication architecture. According to details released regarding the assessment, the testing tool executed the test independently, observed the flaw, and triggered its automated guardrails to halt execution immediately after completing a successful unauthorized login.
Flawed Parameters Allowed Arbitrary Code Delivery
At the center of the flaw was a breakdown in how one-time passcodes (OTPs) were tied to registered customer records. In standard implementations, multi-factor authentication passcodes must strictly be dispatched to contact destinations that are already verified and bound to the customer’s account profile.
However, analysis of the insurer’s login mechanism revealed that the client-side login request passed two distinct parameters controlling code transmission: one defining the delivery channel and another specifying the destination address. Under default operation, the portal delivered authentication codes via SMS to the mobile telephone number already registered to the account.
When the agent intercepted and modified the request, it switched the designated transmission channel from SMS to email and provided an arbitrary email address managed by the testing framework. The insurer’s backend servers processed the incoming request and dispatched the valid authentication code directly to the attacker-specified email address, completely omitting a verification check to ensure that the recipient address belonged to the account owner.
Because the server failed to validate the destination against stored profile records, an external actor possessing basic customer identifiers could potentially route authentication passcodes to an external address and access policyholder accounts without legitimate credentials.
Reconnaissance and Autonomous Execution
The discovery process began during an automated reconnaissance phase, in which the agent inspected client-side JavaScript files and monitored backend API transactions driving the portal’s frontend forms. Through this analysis, the agent detected the request structure governing the OTP delivery mechanism.
Because anomalous or configurable parameters in API requests do not always guarantee an exploitable bug—as backend logic can reject payloads downstream—the agent formulated a hypothesis and constructed a targeted proof-of-concept test against the live production environment. It sent the altered request, retrieved the resulting one-time passcode from its external inbox, and submitted the code to finalize authentication. Once the portal accepted the code and authorized the session, internal safety controls stopped further testing to avoid accessing private customer information, and the finding was formally disclosed to the insurer.
Continuous Testing Versus Point-in-Time Audits
The insurer noted that it had previously undergone a third-party penetration test approximately six months prior to the assessment. However, it remains undetermined precisely when the vulnerable authentication logic was introduced into the production codebase.
Security analysts note that while human penetration testers routinely look for such logical flaws, scheduled point-in-time assessments often struggle to keep pace with continuous software release cycles. In dynamic environments where web applications are updated frequently, vulnerabilities introduced between audit cycles can linger undetected until the next formal testing engagement.
Reflectiz built its agentic framework to address multi-step transactional flows, such as checkout systems and multi-factor authentication gateways, by simulating user journeys within real browser instances. Reflectiz CTO and co-founder Ysrael Gurt, an ethical hacker, designed the system alongside a separate validator agent that independently reproduces suspected findings with identical payloads before issues are reported, reducing false positives while expanding evaluation frequency.
