Microsoft Research described research into a model-modification backdoor on 10 November. According to the company, a model modified in the Lazarus experiment can behave normally until a hidden trigger appears; this makes procurement and permission risks around AI agents tangible for Hungarian businesses.
We do not describe an actual corporate intrusion on the basis of this research demonstration. The subject is a demonstrated attack possibility, not an identified Hungarian victim or incident.
The practical business question is who can connect the procured model to the version actually used. A successful test conversation provides different evidence from documentation of the delivery path and versions. Implementation records can therefore usefully distinguish functional test results from information establishing model provenance.
When reviewing permissions, start with executable actions: what data can the system read or modify? This offers a more concrete basis for decisions than a general label of safety.
We covered language data in our Google Language Explorer introduction and business response processing in our Google Forms feature article. Independent replication of the research and documented evaluation of defensive methods could provide the next substantive development.
Original editorial content. Read our editorial standards.
Budapest Global Review: Microsoft Research examines hidden model modifications. Budapest Global Review, 10 November 2026. https://globalreview.hu/en/cikk/microsoft-lazarus-modell-eredet-2026