Summary
Security researcher Sammy Azdoufal’s report on Energizelab’s Eilik and Eiliko robots describes exposed child profiles, cloud access-control flaws and evidence of third-party access to a shared server. The report says 32 of 40 remotely testable findings remained open at its October 13, 2026 retest.
Security researcher Sammy Azdoufal’s audit of Energizelab’s Eilik and Eiliko companion robots reports weaknesses that could expose children’s conversation records and profile data. The repository says 32 of 40 remotely testable findings were still open at a retest on October 13, 2026, while eight had been fixed.
The report describes a product whose voice conversations depend on cloud services: speech is sent from the robot to Energizelab’s servers, processed by a large language model, then returned as synthesized speech. The researcher says the transcripts of both sides of conversations were stored in clear text, alongside device and network information.
What the audit says it found
The repository describes an investigation involving analysis of Android app versions, live checks of the company’s APIs, firmware reverse engineering across nine product lines, and examination of desktop applications and exposed services. It says the researcher retested findings using non-destructive checks.
Among the reported issues are cloud services using unencrypted HTTP and a file-read flaw that exposed secrets used by the platform. The researcher says those secrets could be used to forge an administrator token accepted by the company’s back-office interface. That interface, according to the report, displays conversation records and child profiles. The researcher’s documented read-path demonstration used their own account; the report says related endpoints lacked ownership checks and could return histories associated with other robots.
The report also describes a nightly pipeline that extracts 24 structured fields from conversations, including a child’s name, age, address, preferences and family relationships. It says the profile count grew from 31,450 on September 1 to more than 32,228 by October 13, 2026.
A separate finding concerns data used to shape the robot’s responses. The researcher says the robot reads a profile stored in Redis as part of its conversation process, and that they changed profile information on researcher-owned devices to alter what the robot treated as true. Those demonstrations used deliberately fictional content.
Reported third-party access and the limits of the tests
The repository says the researcher found webshells and a root-level backdoor on a shared production server, with evidence indicating access by an unauthorized third party. It reports that these remained present at the October retest. The same server was involved in the robot service and firmware infrastructure, according to the researcher.
The report describes a potential route from exposed credentials and weaknesses in the update system to firmware-level access, but explicitly says the component weaknesses were validated separately. It does not claim an end-to-end compromise of the robot fleet or that firmware was deployed to customer devices. The researcher says proof-of-concept scripts and some supporting artifacts are being kept private while the vendor addresses the findings.
The central concern is the combination of a child-facing voice product, cloud-stored conversations and profiles, and reported weaknesses in the systems controlling access to that data. The repository’s status figures describe the infrastructure as of October 13, 2026; remediation after that date is not covered in the supplied report text.